4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * It as also possible to register a callback which is run when the
103 * arc_meta_limit is reached and no buffers can be safely evicted. In
104 * this case the arc user should drop a reference on some arc buffers so
105 * they can be reclaimed and the arc_meta_limit honored. For example,
106 * when using the ZPL each dentry holds a references on a znode. These
107 * dentries must be pruned before the arc buffer holding the znode can
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
125 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
126 * This structure can point either to a block that is still in the cache or to
127 * one that is only accessible in an L2 ARC device, or it can provide
128 * information about a block that was recently evicted. If a block is
129 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
130 * information to retrieve it from the L2ARC device. This information is
131 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
132 * that is in this state cannot access the data directly.
134 * Blocks that are actively being referenced or have not been evicted
135 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
136 * the arc_buf_hdr_t that will point to the data block in memory. A block can
137 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
138 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
139 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
141 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
142 * ability to store the physical data (b_pabd) associated with the DVA of the
143 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
144 * it will match its on-disk compression characteristics. This behavior can be
145 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
146 * compressed ARC functionality is disabled, the b_pabd will point to an
147 * uncompressed version of the on-disk data.
149 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
150 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
151 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
152 * consumer. The ARC will provide references to this data and will keep it
153 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
154 * data block and will evict any arc_buf_t that is no longer referenced. The
155 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
156 * "overhead_size" kstat.
158 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
159 * compressed form. The typical case is that consumers will want uncompressed
160 * data, and when that happens a new data buffer is allocated where the data is
161 * decompressed for them to use. Currently the only consumer who wants
162 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
163 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
164 * with the arc_buf_hdr_t.
166 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
167 * first one is owned by a compressed send consumer (and therefore references
168 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
169 * used by any other consumer (and has its own uncompressed copy of the data
184 * | b_buf +------------>+-----------+ arc_buf_t
185 * | b_pabd +-+ |b_next +---->+-----------+
186 * +-----------+ | |-----------| |b_next +-->NULL
187 * | |b_comp = T | +-----------+
188 * | |b_data +-+ |b_comp = F |
189 * | +-----------+ | |b_data +-+
190 * +->+------+ | +-----------+ |
192 * data | |<--------------+ | uncompressed
193 * +------+ compressed, | data
194 * shared +-->+------+
199 * When a consumer reads a block, the ARC must first look to see if the
200 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
201 * arc_buf_t and either copies uncompressed data into a new data buffer from an
202 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
203 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
204 * hdr is compressed and the desired compression characteristics of the
205 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
206 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
207 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
208 * be anywhere in the hdr's list.
210 * The diagram below shows an example of an uncompressed ARC hdr that is
211 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
212 * the last element in the buf list):
224 * | | arc_buf_t (shared)
225 * | b_buf +------------>+---------+ arc_buf_t
226 * | | |b_next +---->+---------+
227 * | b_pabd +-+ |---------| |b_next +-->NULL
228 * +-----------+ | | | +---------+
230 * | +---------+ | |b_data +-+
231 * +->+------+ | +---------+ |
233 * uncompressed | | | |
236 * | uncompressed | | |
239 * +---------------------------------+
241 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
242 * since the physical block is about to be rewritten. The new data contents
243 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
244 * it may compress the data before writing it to disk. The ARC will be called
245 * with the transformed data and will bcopy the transformed on-disk block into
246 * a newly allocated b_pabd. Writes are always done into buffers which have
247 * either been loaned (and hence are new and don't have other readers) or
248 * buffers which have been released (and hence have their own hdr, if there
249 * were originally other readers of the buf's original hdr). This ensures that
250 * the ARC only needs to update a single buf and its hdr after a write occurs.
252 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
253 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
254 * that when compressed ARC is enabled that the L2ARC blocks are identical
255 * to the on-disk block in the main data pool. This provides a significant
256 * advantage since the ARC can leverage the bp's checksum when reading from the
257 * L2ARC to determine if the contents are valid. However, if the compressed
258 * ARC is disabled, then the L2ARC's block must be transformed to look
259 * like the physical block in the main data pool before comparing the
260 * checksum and determining its validity.
262 * The L1ARC has a slightly different system for storing encrypted data.
263 * Raw (encrypted + possibly compressed) data has a few subtle differences from
264 * data that is just compressed. The biggest difference is that it is not
265 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
266 * The other difference is that encryption cannot be treated as a suggestion.
267 * If a caller would prefer compressed data, but they actually wind up with
268 * uncompressed data the worst thing that could happen is there might be a
269 * performance hit. If the caller requests encrypted data, however, we must be
270 * sure they actually get it or else secret information could be leaked. Raw
271 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
272 * may have both an encrypted version and a decrypted version of its data at
273 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
274 * copied out of this header. To avoid complications with b_pabd, raw buffers
280 #include <sys/spa_impl.h>
281 #include <sys/zio_compress.h>
282 #include <sys/zio_checksum.h>
283 #include <sys/zfs_context.h>
285 #include <sys/refcount.h>
286 #include <sys/vdev.h>
287 #include <sys/vdev_impl.h>
288 #include <sys/dsl_pool.h>
289 #include <sys/zio_checksum.h>
290 #include <sys/multilist.h>
293 #include <sys/fm/fs/zfs.h>
295 #include <sys/shrinker.h>
296 #include <sys/vmsystm.h>
298 #include <linux/page_compat.h>
300 #include <sys/callb.h>
301 #include <sys/kstat.h>
302 #include <sys/dmu_tx.h>
303 #include <zfs_fletcher.h>
304 #include <sys/arc_impl.h>
305 #include <sys/trace_arc.h>
306 #include <sys/aggsum.h>
307 #include <sys/cityhash.h>
310 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
311 boolean_t arc_watch = B_FALSE;
314 static kmutex_t arc_reclaim_lock;
315 static kcondvar_t arc_reclaim_thread_cv;
316 static boolean_t arc_reclaim_thread_exit;
317 static kcondvar_t arc_reclaim_waiters_cv;
320 * The number of headers to evict in arc_evict_state_impl() before
321 * dropping the sublist lock and evicting from another sublist. A lower
322 * value means we're more likely to evict the "correct" header (i.e. the
323 * oldest header in the arc state), but comes with higher overhead
324 * (i.e. more invocations of arc_evict_state_impl()).
326 int zfs_arc_evict_batch_limit = 10;
328 /* number of seconds before growing cache again */
329 static int arc_grow_retry = 5;
331 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
332 int zfs_arc_overflow_shift = 8;
334 /* shift of arc_c for calculating both min and max arc_p */
335 static int arc_p_min_shift = 4;
337 /* log2(fraction of arc to reclaim) */
338 static int arc_shrink_shift = 7;
340 /* percent of pagecache to reclaim arc to */
342 static uint_t zfs_arc_pc_percent = 0;
346 * log2(fraction of ARC which must be free to allow growing).
347 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
348 * when reading a new block into the ARC, we will evict an equal-sized block
351 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
352 * we will still not allow it to grow.
354 int arc_no_grow_shift = 5;
358 * minimum lifespan of a prefetch block in clock ticks
359 * (initialized in arc_init())
361 static int arc_min_prefetch_ms;
362 static int arc_min_prescient_prefetch_ms;
365 * If this percent of memory is free, don't throttle.
367 int arc_lotsfree_percent = 10;
372 * The arc has filled available memory and has now warmed up.
374 static boolean_t arc_warm;
377 * log2 fraction of the zio arena to keep free.
379 int arc_zio_arena_free_shift = 2;
382 * These tunables are for performance analysis.
384 unsigned long zfs_arc_max = 0;
385 unsigned long zfs_arc_min = 0;
386 unsigned long zfs_arc_meta_limit = 0;
387 unsigned long zfs_arc_meta_min = 0;
388 unsigned long zfs_arc_dnode_limit = 0;
389 unsigned long zfs_arc_dnode_reduce_percent = 10;
390 int zfs_arc_grow_retry = 0;
391 int zfs_arc_shrink_shift = 0;
392 int zfs_arc_p_min_shift = 0;
393 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
396 * ARC dirty data constraints for arc_tempreserve_space() throttle.
398 unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
399 unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
400 unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
403 * Enable or disable compressed arc buffers.
405 int zfs_compressed_arc_enabled = B_TRUE;
408 * ARC will evict meta buffers that exceed arc_meta_limit. This
409 * tunable make arc_meta_limit adjustable for different workloads.
411 unsigned long zfs_arc_meta_limit_percent = 75;
414 * Percentage that can be consumed by dnodes of ARC meta buffers.
416 unsigned long zfs_arc_dnode_limit_percent = 10;
419 * These tunables are Linux specific
421 unsigned long zfs_arc_sys_free = 0;
422 int zfs_arc_min_prefetch_ms = 0;
423 int zfs_arc_min_prescient_prefetch_ms = 0;
424 int zfs_arc_p_dampener_disable = 1;
425 int zfs_arc_meta_prune = 10000;
426 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
427 int zfs_arc_meta_adjust_restarts = 4096;
428 int zfs_arc_lotsfree_percent = 10;
431 static arc_state_t ARC_anon;
432 static arc_state_t ARC_mru;
433 static arc_state_t ARC_mru_ghost;
434 static arc_state_t ARC_mfu;
435 static arc_state_t ARC_mfu_ghost;
436 static arc_state_t ARC_l2c_only;
438 typedef struct arc_stats {
439 kstat_named_t arcstat_hits;
440 kstat_named_t arcstat_misses;
441 kstat_named_t arcstat_demand_data_hits;
442 kstat_named_t arcstat_demand_data_misses;
443 kstat_named_t arcstat_demand_metadata_hits;
444 kstat_named_t arcstat_demand_metadata_misses;
445 kstat_named_t arcstat_prefetch_data_hits;
446 kstat_named_t arcstat_prefetch_data_misses;
447 kstat_named_t arcstat_prefetch_metadata_hits;
448 kstat_named_t arcstat_prefetch_metadata_misses;
449 kstat_named_t arcstat_mru_hits;
450 kstat_named_t arcstat_mru_ghost_hits;
451 kstat_named_t arcstat_mfu_hits;
452 kstat_named_t arcstat_mfu_ghost_hits;
453 kstat_named_t arcstat_deleted;
455 * Number of buffers that could not be evicted because the hash lock
456 * was held by another thread. The lock may not necessarily be held
457 * by something using the same buffer, since hash locks are shared
458 * by multiple buffers.
460 kstat_named_t arcstat_mutex_miss;
462 * Number of buffers skipped when updating the access state due to the
463 * header having already been released after acquiring the hash lock.
465 kstat_named_t arcstat_access_skip;
467 * Number of buffers skipped because they have I/O in progress, are
468 * indirect prefetch buffers that have not lived long enough, or are
469 * not from the spa we're trying to evict from.
471 kstat_named_t arcstat_evict_skip;
473 * Number of times arc_evict_state() was unable to evict enough
474 * buffers to reach its target amount.
476 kstat_named_t arcstat_evict_not_enough;
477 kstat_named_t arcstat_evict_l2_cached;
478 kstat_named_t arcstat_evict_l2_eligible;
479 kstat_named_t arcstat_evict_l2_ineligible;
480 kstat_named_t arcstat_evict_l2_skip;
481 kstat_named_t arcstat_hash_elements;
482 kstat_named_t arcstat_hash_elements_max;
483 kstat_named_t arcstat_hash_collisions;
484 kstat_named_t arcstat_hash_chains;
485 kstat_named_t arcstat_hash_chain_max;
486 kstat_named_t arcstat_p;
487 kstat_named_t arcstat_c;
488 kstat_named_t arcstat_c_min;
489 kstat_named_t arcstat_c_max;
490 /* Not updated directly; only synced in arc_kstat_update. */
491 kstat_named_t arcstat_size;
493 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
494 * Note that the compressed bytes may match the uncompressed bytes
495 * if the block is either not compressed or compressed arc is disabled.
497 kstat_named_t arcstat_compressed_size;
499 * Uncompressed size of the data stored in b_pabd. If compressed
500 * arc is disabled then this value will be identical to the stat
503 kstat_named_t arcstat_uncompressed_size;
505 * Number of bytes stored in all the arc_buf_t's. This is classified
506 * as "overhead" since this data is typically short-lived and will
507 * be evicted from the arc when it becomes unreferenced unless the
508 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
509 * values have been set (see comment in dbuf.c for more information).
511 kstat_named_t arcstat_overhead_size;
513 * Number of bytes consumed by internal ARC structures necessary
514 * for tracking purposes; these structures are not actually
515 * backed by ARC buffers. This includes arc_buf_hdr_t structures
516 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
517 * caches), and arc_buf_t structures (allocated via arc_buf_t
519 * Not updated directly; only synced in arc_kstat_update.
521 kstat_named_t arcstat_hdr_size;
523 * Number of bytes consumed by ARC buffers of type equal to
524 * ARC_BUFC_DATA. This is generally consumed by buffers backing
525 * on disk user data (e.g. plain file contents).
526 * Not updated directly; only synced in arc_kstat_update.
528 kstat_named_t arcstat_data_size;
530 * Number of bytes consumed by ARC buffers of type equal to
531 * ARC_BUFC_METADATA. This is generally consumed by buffers
532 * backing on disk data that is used for internal ZFS
533 * structures (e.g. ZAP, dnode, indirect blocks, etc).
534 * Not updated directly; only synced in arc_kstat_update.
536 kstat_named_t arcstat_metadata_size;
538 * Number of bytes consumed by dmu_buf_impl_t objects.
539 * Not updated directly; only synced in arc_kstat_update.
541 kstat_named_t arcstat_dbuf_size;
543 * Number of bytes consumed by dnode_t objects.
544 * Not updated directly; only synced in arc_kstat_update.
546 kstat_named_t arcstat_dnode_size;
548 * Number of bytes consumed by bonus buffers.
549 * Not updated directly; only synced in arc_kstat_update.
551 kstat_named_t arcstat_bonus_size;
553 * Total number of bytes consumed by ARC buffers residing in the
554 * arc_anon state. This includes *all* buffers in the arc_anon
555 * state; e.g. data, metadata, evictable, and unevictable buffers
556 * are all included in this value.
557 * Not updated directly; only synced in arc_kstat_update.
559 kstat_named_t arcstat_anon_size;
561 * Number of bytes consumed by ARC buffers that meet the
562 * following criteria: backing buffers of type ARC_BUFC_DATA,
563 * residing in the arc_anon state, and are eligible for eviction
564 * (e.g. have no outstanding holds on the buffer).
565 * Not updated directly; only synced in arc_kstat_update.
567 kstat_named_t arcstat_anon_evictable_data;
569 * Number of bytes consumed by ARC buffers that meet the
570 * following criteria: backing buffers of type ARC_BUFC_METADATA,
571 * residing in the arc_anon state, and are eligible for eviction
572 * (e.g. have no outstanding holds on the buffer).
573 * Not updated directly; only synced in arc_kstat_update.
575 kstat_named_t arcstat_anon_evictable_metadata;
577 * Total number of bytes consumed by ARC buffers residing in the
578 * arc_mru state. This includes *all* buffers in the arc_mru
579 * state; e.g. data, metadata, evictable, and unevictable buffers
580 * are all included in this value.
581 * Not updated directly; only synced in arc_kstat_update.
583 kstat_named_t arcstat_mru_size;
585 * Number of bytes consumed by ARC buffers that meet the
586 * following criteria: backing buffers of type ARC_BUFC_DATA,
587 * residing in the arc_mru state, and are eligible for eviction
588 * (e.g. have no outstanding holds on the buffer).
589 * Not updated directly; only synced in arc_kstat_update.
591 kstat_named_t arcstat_mru_evictable_data;
593 * Number of bytes consumed by ARC buffers that meet the
594 * following criteria: backing buffers of type ARC_BUFC_METADATA,
595 * residing in the arc_mru state, and are eligible for eviction
596 * (e.g. have no outstanding holds on the buffer).
597 * Not updated directly; only synced in arc_kstat_update.
599 kstat_named_t arcstat_mru_evictable_metadata;
601 * Total number of bytes that *would have been* consumed by ARC
602 * buffers in the arc_mru_ghost state. The key thing to note
603 * here, is the fact that this size doesn't actually indicate
604 * RAM consumption. The ghost lists only consist of headers and
605 * don't actually have ARC buffers linked off of these headers.
606 * Thus, *if* the headers had associated ARC buffers, these
607 * buffers *would have* consumed this number of bytes.
608 * Not updated directly; only synced in arc_kstat_update.
610 kstat_named_t arcstat_mru_ghost_size;
612 * Number of bytes that *would have been* consumed by ARC
613 * buffers that are eligible for eviction, of type
614 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
615 * Not updated directly; only synced in arc_kstat_update.
617 kstat_named_t arcstat_mru_ghost_evictable_data;
619 * Number of bytes that *would have been* consumed by ARC
620 * buffers that are eligible for eviction, of type
621 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
622 * Not updated directly; only synced in arc_kstat_update.
624 kstat_named_t arcstat_mru_ghost_evictable_metadata;
626 * Total number of bytes consumed by ARC buffers residing in the
627 * arc_mfu state. This includes *all* buffers in the arc_mfu
628 * state; e.g. data, metadata, evictable, and unevictable buffers
629 * are all included in this value.
630 * Not updated directly; only synced in arc_kstat_update.
632 kstat_named_t arcstat_mfu_size;
634 * Number of bytes consumed by ARC buffers that are eligible for
635 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
637 * Not updated directly; only synced in arc_kstat_update.
639 kstat_named_t arcstat_mfu_evictable_data;
641 * Number of bytes consumed by ARC buffers that are eligible for
642 * eviction, of type ARC_BUFC_METADATA, and reside in the
644 * Not updated directly; only synced in arc_kstat_update.
646 kstat_named_t arcstat_mfu_evictable_metadata;
648 * Total number of bytes that *would have been* consumed by ARC
649 * buffers in the arc_mfu_ghost state. See the comment above
650 * arcstat_mru_ghost_size for more details.
651 * Not updated directly; only synced in arc_kstat_update.
653 kstat_named_t arcstat_mfu_ghost_size;
655 * Number of bytes that *would have been* consumed by ARC
656 * buffers that are eligible for eviction, of type
657 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
658 * Not updated directly; only synced in arc_kstat_update.
660 kstat_named_t arcstat_mfu_ghost_evictable_data;
662 * Number of bytes that *would have been* consumed by ARC
663 * buffers that are eligible for eviction, of type
664 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
665 * Not updated directly; only synced in arc_kstat_update.
667 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
668 kstat_named_t arcstat_l2_hits;
669 kstat_named_t arcstat_l2_misses;
670 kstat_named_t arcstat_l2_feeds;
671 kstat_named_t arcstat_l2_rw_clash;
672 kstat_named_t arcstat_l2_read_bytes;
673 kstat_named_t arcstat_l2_write_bytes;
674 kstat_named_t arcstat_l2_writes_sent;
675 kstat_named_t arcstat_l2_writes_done;
676 kstat_named_t arcstat_l2_writes_error;
677 kstat_named_t arcstat_l2_writes_lock_retry;
678 kstat_named_t arcstat_l2_evict_lock_retry;
679 kstat_named_t arcstat_l2_evict_reading;
680 kstat_named_t arcstat_l2_evict_l1cached;
681 kstat_named_t arcstat_l2_free_on_write;
682 kstat_named_t arcstat_l2_abort_lowmem;
683 kstat_named_t arcstat_l2_cksum_bad;
684 kstat_named_t arcstat_l2_io_error;
685 kstat_named_t arcstat_l2_lsize;
686 kstat_named_t arcstat_l2_psize;
687 /* Not updated directly; only synced in arc_kstat_update. */
688 kstat_named_t arcstat_l2_hdr_size;
689 kstat_named_t arcstat_memory_throttle_count;
690 kstat_named_t arcstat_memory_direct_count;
691 kstat_named_t arcstat_memory_indirect_count;
692 kstat_named_t arcstat_memory_all_bytes;
693 kstat_named_t arcstat_memory_free_bytes;
694 kstat_named_t arcstat_memory_available_bytes;
695 kstat_named_t arcstat_no_grow;
696 kstat_named_t arcstat_tempreserve;
697 kstat_named_t arcstat_loaned_bytes;
698 kstat_named_t arcstat_prune;
699 /* Not updated directly; only synced in arc_kstat_update. */
700 kstat_named_t arcstat_meta_used;
701 kstat_named_t arcstat_meta_limit;
702 kstat_named_t arcstat_dnode_limit;
703 kstat_named_t arcstat_meta_max;
704 kstat_named_t arcstat_meta_min;
705 kstat_named_t arcstat_async_upgrade_sync;
706 kstat_named_t arcstat_demand_hit_predictive_prefetch;
707 kstat_named_t arcstat_demand_hit_prescient_prefetch;
708 kstat_named_t arcstat_need_free;
709 kstat_named_t arcstat_sys_free;
710 kstat_named_t arcstat_raw_size;
713 static arc_stats_t arc_stats = {
714 { "hits", KSTAT_DATA_UINT64 },
715 { "misses", KSTAT_DATA_UINT64 },
716 { "demand_data_hits", KSTAT_DATA_UINT64 },
717 { "demand_data_misses", KSTAT_DATA_UINT64 },
718 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
719 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
720 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
721 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
722 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
723 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
724 { "mru_hits", KSTAT_DATA_UINT64 },
725 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
726 { "mfu_hits", KSTAT_DATA_UINT64 },
727 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
728 { "deleted", KSTAT_DATA_UINT64 },
729 { "mutex_miss", KSTAT_DATA_UINT64 },
730 { "access_skip", KSTAT_DATA_UINT64 },
731 { "evict_skip", KSTAT_DATA_UINT64 },
732 { "evict_not_enough", KSTAT_DATA_UINT64 },
733 { "evict_l2_cached", KSTAT_DATA_UINT64 },
734 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
735 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
736 { "evict_l2_skip", KSTAT_DATA_UINT64 },
737 { "hash_elements", KSTAT_DATA_UINT64 },
738 { "hash_elements_max", KSTAT_DATA_UINT64 },
739 { "hash_collisions", KSTAT_DATA_UINT64 },
740 { "hash_chains", KSTAT_DATA_UINT64 },
741 { "hash_chain_max", KSTAT_DATA_UINT64 },
742 { "p", KSTAT_DATA_UINT64 },
743 { "c", KSTAT_DATA_UINT64 },
744 { "c_min", KSTAT_DATA_UINT64 },
745 { "c_max", KSTAT_DATA_UINT64 },
746 { "size", KSTAT_DATA_UINT64 },
747 { "compressed_size", KSTAT_DATA_UINT64 },
748 { "uncompressed_size", KSTAT_DATA_UINT64 },
749 { "overhead_size", KSTAT_DATA_UINT64 },
750 { "hdr_size", KSTAT_DATA_UINT64 },
751 { "data_size", KSTAT_DATA_UINT64 },
752 { "metadata_size", KSTAT_DATA_UINT64 },
753 { "dbuf_size", KSTAT_DATA_UINT64 },
754 { "dnode_size", KSTAT_DATA_UINT64 },
755 { "bonus_size", KSTAT_DATA_UINT64 },
756 { "anon_size", KSTAT_DATA_UINT64 },
757 { "anon_evictable_data", KSTAT_DATA_UINT64 },
758 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
759 { "mru_size", KSTAT_DATA_UINT64 },
760 { "mru_evictable_data", KSTAT_DATA_UINT64 },
761 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
762 { "mru_ghost_size", KSTAT_DATA_UINT64 },
763 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
764 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
765 { "mfu_size", KSTAT_DATA_UINT64 },
766 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
767 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
768 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
769 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
770 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
771 { "l2_hits", KSTAT_DATA_UINT64 },
772 { "l2_misses", KSTAT_DATA_UINT64 },
773 { "l2_feeds", KSTAT_DATA_UINT64 },
774 { "l2_rw_clash", KSTAT_DATA_UINT64 },
775 { "l2_read_bytes", KSTAT_DATA_UINT64 },
776 { "l2_write_bytes", KSTAT_DATA_UINT64 },
777 { "l2_writes_sent", KSTAT_DATA_UINT64 },
778 { "l2_writes_done", KSTAT_DATA_UINT64 },
779 { "l2_writes_error", KSTAT_DATA_UINT64 },
780 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
781 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
782 { "l2_evict_reading", KSTAT_DATA_UINT64 },
783 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
784 { "l2_free_on_write", KSTAT_DATA_UINT64 },
785 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
786 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
787 { "l2_io_error", KSTAT_DATA_UINT64 },
788 { "l2_size", KSTAT_DATA_UINT64 },
789 { "l2_asize", KSTAT_DATA_UINT64 },
790 { "l2_hdr_size", KSTAT_DATA_UINT64 },
791 { "memory_throttle_count", KSTAT_DATA_UINT64 },
792 { "memory_direct_count", KSTAT_DATA_UINT64 },
793 { "memory_indirect_count", KSTAT_DATA_UINT64 },
794 { "memory_all_bytes", KSTAT_DATA_UINT64 },
795 { "memory_free_bytes", KSTAT_DATA_UINT64 },
796 { "memory_available_bytes", KSTAT_DATA_INT64 },
797 { "arc_no_grow", KSTAT_DATA_UINT64 },
798 { "arc_tempreserve", KSTAT_DATA_UINT64 },
799 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
800 { "arc_prune", KSTAT_DATA_UINT64 },
801 { "arc_meta_used", KSTAT_DATA_UINT64 },
802 { "arc_meta_limit", KSTAT_DATA_UINT64 },
803 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
804 { "arc_meta_max", KSTAT_DATA_UINT64 },
805 { "arc_meta_min", KSTAT_DATA_UINT64 },
806 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
807 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
808 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
809 { "arc_need_free", KSTAT_DATA_UINT64 },
810 { "arc_sys_free", KSTAT_DATA_UINT64 },
811 { "arc_raw_size", KSTAT_DATA_UINT64 }
814 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
816 #define ARCSTAT_INCR(stat, val) \
817 atomic_add_64(&arc_stats.stat.value.ui64, (val))
819 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
820 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
822 #define ARCSTAT_MAX(stat, val) { \
824 while ((val) > (m = arc_stats.stat.value.ui64) && \
825 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
829 #define ARCSTAT_MAXSTAT(stat) \
830 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
833 * We define a macro to allow ARC hits/misses to be easily broken down by
834 * two separate conditions, giving a total of four different subtypes for
835 * each of hits and misses (so eight statistics total).
837 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
840 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
842 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
846 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
848 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
853 static arc_state_t *arc_anon;
854 static arc_state_t *arc_mru;
855 static arc_state_t *arc_mru_ghost;
856 static arc_state_t *arc_mfu;
857 static arc_state_t *arc_mfu_ghost;
858 static arc_state_t *arc_l2c_only;
861 * There are several ARC variables that are critical to export as kstats --
862 * but we don't want to have to grovel around in the kstat whenever we wish to
863 * manipulate them. For these variables, we therefore define them to be in
864 * terms of the statistic variable. This assures that we are not introducing
865 * the possibility of inconsistency by having shadow copies of the variables,
866 * while still allowing the code to be readable.
868 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
869 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
870 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
871 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
872 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
873 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
874 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
875 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
876 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
877 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
878 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
879 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
880 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
882 /* size of all b_rabd's in entire arc */
883 #define arc_raw_size ARCSTAT(arcstat_raw_size)
884 /* compressed size of entire arc */
885 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
886 /* uncompressed size of entire arc */
887 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
888 /* number of bytes in the arc from arc_buf_t's */
889 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
892 * There are also some ARC variables that we want to export, but that are
893 * updated so often that having the canonical representation be the statistic
894 * variable causes a performance bottleneck. We want to use aggsum_t's for these
895 * instead, but still be able to export the kstat in the same way as before.
896 * The solution is to always use the aggsum version, except in the kstat update
900 aggsum_t arc_meta_used;
901 aggsum_t astat_data_size;
902 aggsum_t astat_metadata_size;
903 aggsum_t astat_dbuf_size;
904 aggsum_t astat_dnode_size;
905 aggsum_t astat_bonus_size;
906 aggsum_t astat_hdr_size;
907 aggsum_t astat_l2_hdr_size;
909 static list_t arc_prune_list;
910 static kmutex_t arc_prune_mtx;
911 static taskq_t *arc_prune_taskq;
913 #define GHOST_STATE(state) \
914 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
915 (state) == arc_l2c_only)
917 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
918 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
919 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
920 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
921 #define HDR_PRESCIENT_PREFETCH(hdr) \
922 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
923 #define HDR_COMPRESSION_ENABLED(hdr) \
924 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
926 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
927 #define HDR_L2_READING(hdr) \
928 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
929 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
930 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
931 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
932 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
933 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
934 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
935 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
937 #define HDR_ISTYPE_METADATA(hdr) \
938 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
939 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
941 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
942 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
943 #define HDR_HAS_RABD(hdr) \
944 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
945 (hdr)->b_crypt_hdr.b_rabd != NULL)
946 #define HDR_ENCRYPTED(hdr) \
947 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
948 #define HDR_AUTHENTICATED(hdr) \
949 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
951 /* For storing compression mode in b_flags */
952 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
954 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
955 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
956 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
957 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
959 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
960 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
961 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
962 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
968 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
969 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
970 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
973 * Hash table routines
976 #define HT_LOCK_ALIGN 64
977 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
982 unsigned char pad[HT_LOCK_PAD];
986 #define BUF_LOCKS 8192
987 typedef struct buf_hash_table {
989 arc_buf_hdr_t **ht_table;
990 struct ht_lock ht_locks[BUF_LOCKS];
993 static buf_hash_table_t buf_hash_table;
995 #define BUF_HASH_INDEX(spa, dva, birth) \
996 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
997 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
998 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
999 #define HDR_LOCK(hdr) \
1000 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1002 uint64_t zfs_crc64_table[256];
1008 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1009 #define L2ARC_HEADROOM 2 /* num of writes */
1012 * If we discover during ARC scan any buffers to be compressed, we boost
1013 * our headroom for the next scanning cycle by this percentage multiple.
1015 #define L2ARC_HEADROOM_BOOST 200
1016 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1017 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1020 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
1021 * and each of the state has two types: data and metadata.
1023 #define L2ARC_FEED_TYPES 4
1025 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1026 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1028 /* L2ARC Performance Tunables */
1029 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
1030 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
1031 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
1032 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1033 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1034 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
1035 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1036 int l2arc_feed_again = B_TRUE; /* turbo warmup */
1037 int l2arc_norw = B_FALSE; /* no reads during writes */
1042 static list_t L2ARC_dev_list; /* device list */
1043 static list_t *l2arc_dev_list; /* device list pointer */
1044 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1045 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1046 static list_t L2ARC_free_on_write; /* free after write buf list */
1047 static list_t *l2arc_free_on_write; /* free after write list ptr */
1048 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1049 static uint64_t l2arc_ndev; /* number of devices */
1051 typedef struct l2arc_read_callback {
1052 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1053 blkptr_t l2rcb_bp; /* original blkptr */
1054 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1055 int l2rcb_flags; /* original flags */
1056 abd_t *l2rcb_abd; /* temporary buffer */
1057 } l2arc_read_callback_t;
1059 typedef struct l2arc_data_free {
1060 /* protected by l2arc_free_on_write_mtx */
1063 arc_buf_contents_t l2df_type;
1064 list_node_t l2df_list_node;
1065 } l2arc_data_free_t;
1067 typedef enum arc_fill_flags {
1068 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
1069 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
1070 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
1071 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
1072 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
1075 static kmutex_t l2arc_feed_thr_lock;
1076 static kcondvar_t l2arc_feed_thr_cv;
1077 static uint8_t l2arc_thread_exit;
1079 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1080 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1081 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1082 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1083 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1084 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1085 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
1086 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, boolean_t);
1087 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1088 static boolean_t arc_is_overflowing(void);
1089 static void arc_buf_watch(arc_buf_t *);
1090 static void arc_tuning_update(void);
1091 static void arc_prune_async(int64_t);
1092 static uint64_t arc_all_memory(void);
1094 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1095 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1096 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1097 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1099 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1100 static void l2arc_read_done(zio_t *);
1104 * We use Cityhash for this. It's fast, and has good hash properties without
1105 * requiring any large static buffers.
1108 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1110 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1113 #define HDR_EMPTY(hdr) \
1114 ((hdr)->b_dva.dva_word[0] == 0 && \
1115 (hdr)->b_dva.dva_word[1] == 0)
1117 #define HDR_EQUAL(spa, dva, birth, hdr) \
1118 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1119 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1120 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1123 buf_discard_identity(arc_buf_hdr_t *hdr)
1125 hdr->b_dva.dva_word[0] = 0;
1126 hdr->b_dva.dva_word[1] = 0;
1130 static arc_buf_hdr_t *
1131 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1133 const dva_t *dva = BP_IDENTITY(bp);
1134 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1135 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1136 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1139 mutex_enter(hash_lock);
1140 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1141 hdr = hdr->b_hash_next) {
1142 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1147 mutex_exit(hash_lock);
1153 * Insert an entry into the hash table. If there is already an element
1154 * equal to elem in the hash table, then the already existing element
1155 * will be returned and the new element will not be inserted.
1156 * Otherwise returns NULL.
1157 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1159 static arc_buf_hdr_t *
1160 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1162 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1163 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1164 arc_buf_hdr_t *fhdr;
1167 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1168 ASSERT(hdr->b_birth != 0);
1169 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1171 if (lockp != NULL) {
1173 mutex_enter(hash_lock);
1175 ASSERT(MUTEX_HELD(hash_lock));
1178 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1179 fhdr = fhdr->b_hash_next, i++) {
1180 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1184 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1185 buf_hash_table.ht_table[idx] = hdr;
1186 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1188 /* collect some hash table performance data */
1190 ARCSTAT_BUMP(arcstat_hash_collisions);
1192 ARCSTAT_BUMP(arcstat_hash_chains);
1194 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1197 ARCSTAT_BUMP(arcstat_hash_elements);
1198 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1204 buf_hash_remove(arc_buf_hdr_t *hdr)
1206 arc_buf_hdr_t *fhdr, **hdrp;
1207 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1209 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1210 ASSERT(HDR_IN_HASH_TABLE(hdr));
1212 hdrp = &buf_hash_table.ht_table[idx];
1213 while ((fhdr = *hdrp) != hdr) {
1214 ASSERT3P(fhdr, !=, NULL);
1215 hdrp = &fhdr->b_hash_next;
1217 *hdrp = hdr->b_hash_next;
1218 hdr->b_hash_next = NULL;
1219 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1221 /* collect some hash table performance data */
1222 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1224 if (buf_hash_table.ht_table[idx] &&
1225 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1226 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1230 * Global data structures and functions for the buf kmem cache.
1233 static kmem_cache_t *hdr_full_cache;
1234 static kmem_cache_t *hdr_full_crypt_cache;
1235 static kmem_cache_t *hdr_l2only_cache;
1236 static kmem_cache_t *buf_cache;
1243 #if defined(_KERNEL)
1245 * Large allocations which do not require contiguous pages
1246 * should be using vmem_free() in the linux kernel\
1248 vmem_free(buf_hash_table.ht_table,
1249 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1251 kmem_free(buf_hash_table.ht_table,
1252 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1254 for (i = 0; i < BUF_LOCKS; i++)
1255 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1256 kmem_cache_destroy(hdr_full_cache);
1257 kmem_cache_destroy(hdr_full_crypt_cache);
1258 kmem_cache_destroy(hdr_l2only_cache);
1259 kmem_cache_destroy(buf_cache);
1263 * Constructor callback - called when the cache is empty
1264 * and a new buf is requested.
1268 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1270 arc_buf_hdr_t *hdr = vbuf;
1272 bzero(hdr, HDR_FULL_SIZE);
1273 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1274 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1275 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1276 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1277 list_link_init(&hdr->b_l1hdr.b_arc_node);
1278 list_link_init(&hdr->b_l2hdr.b_l2node);
1279 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1280 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1287 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1289 arc_buf_hdr_t *hdr = vbuf;
1291 hdr_full_cons(vbuf, unused, kmflag);
1292 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1293 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1300 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1302 arc_buf_hdr_t *hdr = vbuf;
1304 bzero(hdr, HDR_L2ONLY_SIZE);
1305 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1312 buf_cons(void *vbuf, void *unused, int kmflag)
1314 arc_buf_t *buf = vbuf;
1316 bzero(buf, sizeof (arc_buf_t));
1317 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1318 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1324 * Destructor callback - called when a cached buf is
1325 * no longer required.
1329 hdr_full_dest(void *vbuf, void *unused)
1331 arc_buf_hdr_t *hdr = vbuf;
1333 ASSERT(HDR_EMPTY(hdr));
1334 cv_destroy(&hdr->b_l1hdr.b_cv);
1335 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1336 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1337 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1338 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1343 hdr_full_crypt_dest(void *vbuf, void *unused)
1345 arc_buf_hdr_t *hdr = vbuf;
1347 hdr_full_dest(vbuf, unused);
1348 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1353 hdr_l2only_dest(void *vbuf, void *unused)
1355 ASSERTV(arc_buf_hdr_t *hdr = vbuf);
1357 ASSERT(HDR_EMPTY(hdr));
1358 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1363 buf_dest(void *vbuf, void *unused)
1365 arc_buf_t *buf = vbuf;
1367 mutex_destroy(&buf->b_evict_lock);
1368 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1372 * Reclaim callback -- invoked when memory is low.
1376 hdr_recl(void *unused)
1378 dprintf("hdr_recl called\n");
1380 * umem calls the reclaim func when we destroy the buf cache,
1381 * which is after we do arc_fini().
1384 cv_signal(&arc_reclaim_thread_cv);
1390 uint64_t *ct = NULL;
1391 uint64_t hsize = 1ULL << 12;
1395 * The hash table is big enough to fill all of physical memory
1396 * with an average block size of zfs_arc_average_blocksize (default 8K).
1397 * By default, the table will take up
1398 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1400 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1403 buf_hash_table.ht_mask = hsize - 1;
1404 #if defined(_KERNEL)
1406 * Large allocations which do not require contiguous pages
1407 * should be using vmem_alloc() in the linux kernel
1409 buf_hash_table.ht_table =
1410 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1412 buf_hash_table.ht_table =
1413 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1415 if (buf_hash_table.ht_table == NULL) {
1416 ASSERT(hsize > (1ULL << 8));
1421 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1422 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1423 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1424 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1425 hdr_recl, NULL, NULL, 0);
1426 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1427 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1429 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1430 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1432 for (i = 0; i < 256; i++)
1433 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1434 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1436 for (i = 0; i < BUF_LOCKS; i++) {
1437 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1438 NULL, MUTEX_DEFAULT, NULL);
1442 #define ARC_MINTIME (hz>>4) /* 62 ms */
1445 * This is the size that the buf occupies in memory. If the buf is compressed,
1446 * it will correspond to the compressed size. You should use this method of
1447 * getting the buf size unless you explicitly need the logical size.
1450 arc_buf_size(arc_buf_t *buf)
1452 return (ARC_BUF_COMPRESSED(buf) ?
1453 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1457 arc_buf_lsize(arc_buf_t *buf)
1459 return (HDR_GET_LSIZE(buf->b_hdr));
1463 * This function will return B_TRUE if the buffer is encrypted in memory.
1464 * This buffer can be decrypted by calling arc_untransform().
1467 arc_is_encrypted(arc_buf_t *buf)
1469 return (ARC_BUF_ENCRYPTED(buf) != 0);
1473 * Returns B_TRUE if the buffer represents data that has not had its MAC
1477 arc_is_unauthenticated(arc_buf_t *buf)
1479 return (HDR_NOAUTH(buf->b_hdr) != 0);
1483 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1484 uint8_t *iv, uint8_t *mac)
1486 arc_buf_hdr_t *hdr = buf->b_hdr;
1488 ASSERT(HDR_PROTECTED(hdr));
1490 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1491 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1492 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1493 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1494 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1498 * Indicates how this buffer is compressed in memory. If it is not compressed
1499 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1500 * arc_untransform() as long as it is also unencrypted.
1503 arc_get_compression(arc_buf_t *buf)
1505 return (ARC_BUF_COMPRESSED(buf) ?
1506 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1510 * Return the compression algorithm used to store this data in the ARC. If ARC
1511 * compression is enabled or this is an encrypted block, this will be the same
1512 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1514 static inline enum zio_compress
1515 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1517 return (HDR_COMPRESSION_ENABLED(hdr) ?
1518 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1521 static inline boolean_t
1522 arc_buf_is_shared(arc_buf_t *buf)
1524 boolean_t shared = (buf->b_data != NULL &&
1525 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1526 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1527 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1528 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1529 IMPLY(shared, ARC_BUF_SHARED(buf));
1530 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1533 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1534 * already being shared" requirement prevents us from doing that.
1541 * Free the checksum associated with this header. If there is no checksum, this
1545 arc_cksum_free(arc_buf_hdr_t *hdr)
1547 ASSERT(HDR_HAS_L1HDR(hdr));
1549 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1550 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1551 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1552 hdr->b_l1hdr.b_freeze_cksum = NULL;
1554 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1558 * Return true iff at least one of the bufs on hdr is not compressed.
1559 * Encrypted buffers count as compressed.
1562 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1564 ASSERT(hdr->b_l1hdr.b_state == arc_anon ||
1565 MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1567 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1568 if (!ARC_BUF_COMPRESSED(b)) {
1577 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1578 * matches the checksum that is stored in the hdr. If there is no checksum,
1579 * or if the buf is compressed, this is a no-op.
1582 arc_cksum_verify(arc_buf_t *buf)
1584 arc_buf_hdr_t *hdr = buf->b_hdr;
1587 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1590 if (ARC_BUF_COMPRESSED(buf))
1593 ASSERT(HDR_HAS_L1HDR(hdr));
1595 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1597 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1598 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1602 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1603 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1604 panic("buffer modified while frozen!");
1605 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1609 * This function makes the assumption that data stored in the L2ARC
1610 * will be transformed exactly as it is in the main pool. Because of
1611 * this we can verify the checksum against the reading process's bp.
1614 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1616 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1617 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1620 * Block pointers always store the checksum for the logical data.
1621 * If the block pointer has the gang bit set, then the checksum
1622 * it represents is for the reconstituted data and not for an
1623 * individual gang member. The zio pipeline, however, must be able to
1624 * determine the checksum of each of the gang constituents so it
1625 * treats the checksum comparison differently than what we need
1626 * for l2arc blocks. This prevents us from using the
1627 * zio_checksum_error() interface directly. Instead we must call the
1628 * zio_checksum_error_impl() so that we can ensure the checksum is
1629 * generated using the correct checksum algorithm and accounts for the
1630 * logical I/O size and not just a gang fragment.
1632 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1633 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1634 zio->io_offset, NULL) == 0);
1638 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1639 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1640 * isn't modified later on. If buf is compressed or there is already a checksum
1641 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1644 arc_cksum_compute(arc_buf_t *buf)
1646 arc_buf_hdr_t *hdr = buf->b_hdr;
1648 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1651 ASSERT(HDR_HAS_L1HDR(hdr));
1653 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1654 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1655 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1659 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1660 ASSERT(!ARC_BUF_COMPRESSED(buf));
1661 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1663 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1664 hdr->b_l1hdr.b_freeze_cksum);
1665 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1671 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1673 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1679 arc_buf_unwatch(arc_buf_t *buf)
1683 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1684 PROT_READ | PROT_WRITE));
1691 arc_buf_watch(arc_buf_t *buf)
1695 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1700 static arc_buf_contents_t
1701 arc_buf_type(arc_buf_hdr_t *hdr)
1703 arc_buf_contents_t type;
1704 if (HDR_ISTYPE_METADATA(hdr)) {
1705 type = ARC_BUFC_METADATA;
1707 type = ARC_BUFC_DATA;
1709 VERIFY3U(hdr->b_type, ==, type);
1714 arc_is_metadata(arc_buf_t *buf)
1716 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1720 arc_bufc_to_flags(arc_buf_contents_t type)
1724 /* metadata field is 0 if buffer contains normal data */
1726 case ARC_BUFC_METADATA:
1727 return (ARC_FLAG_BUFC_METADATA);
1731 panic("undefined ARC buffer type!");
1732 return ((uint32_t)-1);
1736 arc_buf_thaw(arc_buf_t *buf)
1738 arc_buf_hdr_t *hdr = buf->b_hdr;
1740 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1741 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1743 arc_cksum_verify(buf);
1746 * Compressed buffers do not manipulate the b_freeze_cksum.
1748 if (ARC_BUF_COMPRESSED(buf))
1751 ASSERT(HDR_HAS_L1HDR(hdr));
1752 arc_cksum_free(hdr);
1753 arc_buf_unwatch(buf);
1757 arc_buf_freeze(arc_buf_t *buf)
1759 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1762 if (ARC_BUF_COMPRESSED(buf))
1765 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1766 arc_cksum_compute(buf);
1770 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1771 * the following functions should be used to ensure that the flags are
1772 * updated in a thread-safe way. When manipulating the flags either
1773 * the hash_lock must be held or the hdr must be undiscoverable. This
1774 * ensures that we're not racing with any other threads when updating
1778 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1780 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1781 hdr->b_flags |= flags;
1785 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1787 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1788 hdr->b_flags &= ~flags;
1792 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1793 * done in a special way since we have to clear and set bits
1794 * at the same time. Consumers that wish to set the compression bits
1795 * must use this function to ensure that the flags are updated in
1796 * thread-safe manner.
1799 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1801 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1804 * Holes and embedded blocks will always have a psize = 0 so
1805 * we ignore the compression of the blkptr and set the
1806 * want to uncompress them. Mark them as uncompressed.
1808 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1809 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1810 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1812 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1813 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1816 HDR_SET_COMPRESS(hdr, cmp);
1817 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1821 * Looks for another buf on the same hdr which has the data decompressed, copies
1822 * from it, and returns true. If no such buf exists, returns false.
1825 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1827 arc_buf_hdr_t *hdr = buf->b_hdr;
1828 boolean_t copied = B_FALSE;
1830 ASSERT(HDR_HAS_L1HDR(hdr));
1831 ASSERT3P(buf->b_data, !=, NULL);
1832 ASSERT(!ARC_BUF_COMPRESSED(buf));
1834 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1835 from = from->b_next) {
1836 /* can't use our own data buffer */
1841 if (!ARC_BUF_COMPRESSED(from)) {
1842 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1849 * There were no decompressed bufs, so there should not be a
1850 * checksum on the hdr either.
1852 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1858 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1861 arc_hdr_size(arc_buf_hdr_t *hdr)
1865 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1866 HDR_GET_PSIZE(hdr) > 0) {
1867 size = HDR_GET_PSIZE(hdr);
1869 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1870 size = HDR_GET_LSIZE(hdr);
1876 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1880 uint64_t lsize = HDR_GET_LSIZE(hdr);
1881 uint64_t psize = HDR_GET_PSIZE(hdr);
1882 void *tmpbuf = NULL;
1883 abd_t *abd = hdr->b_l1hdr.b_pabd;
1885 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1886 ASSERT(HDR_AUTHENTICATED(hdr));
1887 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1890 * The MAC is calculated on the compressed data that is stored on disk.
1891 * However, if compressed arc is disabled we will only have the
1892 * decompressed data available to us now. Compress it into a temporary
1893 * abd so we can verify the MAC. The performance overhead of this will
1894 * be relatively low, since most objects in an encrypted objset will
1895 * be encrypted (instead of authenticated) anyway.
1897 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1898 !HDR_COMPRESSION_ENABLED(hdr)) {
1899 tmpbuf = zio_buf_alloc(lsize);
1900 abd = abd_get_from_buf(tmpbuf, lsize);
1901 abd_take_ownership_of_buf(abd, B_TRUE);
1903 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1904 hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1905 ASSERT3U(csize, <=, psize);
1906 abd_zero_off(abd, csize, psize - csize);
1910 * Authentication is best effort. We authenticate whenever the key is
1911 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1913 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1914 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1915 ASSERT3U(lsize, ==, psize);
1916 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1917 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1919 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1920 hdr->b_crypt_hdr.b_mac);
1924 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1925 else if (ret != ENOENT)
1941 * This function will take a header that only has raw encrypted data in
1942 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1943 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1944 * also decompress the data.
1947 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1952 boolean_t no_crypt = B_FALSE;
1953 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1955 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1956 ASSERT(HDR_ENCRYPTED(hdr));
1958 arc_hdr_alloc_abd(hdr, B_FALSE);
1960 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1961 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1962 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1963 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1968 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1969 HDR_GET_PSIZE(hdr));
1973 * If this header has disabled arc compression but the b_pabd is
1974 * compressed after decrypting it, we need to decompress the newly
1977 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1978 !HDR_COMPRESSION_ENABLED(hdr)) {
1980 * We want to make sure that we are correctly honoring the
1981 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1982 * and then loan a buffer from it, rather than allocating a
1983 * linear buffer and wrapping it in an abd later.
1985 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
1986 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1988 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1989 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1990 HDR_GET_LSIZE(hdr));
1992 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1996 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1997 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1998 arc_hdr_size(hdr), hdr);
1999 hdr->b_l1hdr.b_pabd = cabd;
2005 arc_hdr_free_abd(hdr, B_FALSE);
2007 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
2013 * This function is called during arc_buf_fill() to prepare the header's
2014 * abd plaintext pointer for use. This involves authenticated protected
2015 * data and decrypting encrypted data into the plaintext abd.
2018 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
2019 const zbookmark_phys_t *zb, boolean_t noauth)
2023 ASSERT(HDR_PROTECTED(hdr));
2025 if (hash_lock != NULL)
2026 mutex_enter(hash_lock);
2028 if (HDR_NOAUTH(hdr) && !noauth) {
2030 * The caller requested authenticated data but our data has
2031 * not been authenticated yet. Verify the MAC now if we can.
2033 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
2036 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
2038 * If we only have the encrypted version of the data, but the
2039 * unencrypted version was requested we take this opportunity
2040 * to store the decrypted version in the header for future use.
2042 ret = arc_hdr_decrypt(hdr, spa, zb);
2047 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2049 if (hash_lock != NULL)
2050 mutex_exit(hash_lock);
2055 if (hash_lock != NULL)
2056 mutex_exit(hash_lock);
2062 * This function is used by the dbuf code to decrypt bonus buffers in place.
2063 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
2064 * block, so we use the hash lock here to protect against concurrent calls to
2068 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
2070 arc_buf_hdr_t *hdr = buf->b_hdr;
2072 ASSERT(HDR_ENCRYPTED(hdr));
2073 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2074 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2075 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2077 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
2079 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
2080 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2081 hdr->b_crypt_hdr.b_ebufcnt -= 1;
2085 * Given a buf that has a data buffer attached to it, this function will
2086 * efficiently fill the buf with data of the specified compression setting from
2087 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2088 * are already sharing a data buf, no copy is performed.
2090 * If the buf is marked as compressed but uncompressed data was requested, this
2091 * will allocate a new data buffer for the buf, remove that flag, and fill the
2092 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2093 * uncompressed data, and (since we haven't added support for it yet) if you
2094 * want compressed data your buf must already be marked as compressed and have
2095 * the correct-sized data buffer.
2098 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2099 arc_fill_flags_t flags)
2102 arc_buf_hdr_t *hdr = buf->b_hdr;
2103 boolean_t hdr_compressed =
2104 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
2105 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2106 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2107 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2108 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2110 ASSERT3P(buf->b_data, !=, NULL);
2111 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2112 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2113 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2114 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2115 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2116 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2119 * If the caller wanted encrypted data we just need to copy it from
2120 * b_rabd and potentially byteswap it. We won't be able to do any
2121 * further transforms on it.
2124 ASSERT(HDR_HAS_RABD(hdr));
2125 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2126 HDR_GET_PSIZE(hdr));
2131 * Adjust encrypted and authenticated headers to accomodate
2132 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2133 * allowed to fail decryption due to keys not being loaded
2134 * without being marked as an IO error.
2136 if (HDR_PROTECTED(hdr)) {
2137 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2138 zb, !!(flags & ARC_FILL_NOAUTH));
2139 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2141 } else if (error != 0) {
2142 if (hash_lock != NULL)
2143 mutex_enter(hash_lock);
2144 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2145 if (hash_lock != NULL)
2146 mutex_exit(hash_lock);
2152 * There is a special case here for dnode blocks which are
2153 * decrypting their bonus buffers. These blocks may request to
2154 * be decrypted in-place. This is necessary because there may
2155 * be many dnodes pointing into this buffer and there is
2156 * currently no method to synchronize replacing the backing
2157 * b_data buffer and updating all of the pointers. Here we use
2158 * the hash lock to ensure there are no races. If the need
2159 * arises for other types to be decrypted in-place, they must
2160 * add handling here as well.
2162 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2163 ASSERT(!hdr_compressed);
2164 ASSERT(!compressed);
2167 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2168 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2170 if (hash_lock != NULL)
2171 mutex_enter(hash_lock);
2172 arc_buf_untransform_in_place(buf, hash_lock);
2173 if (hash_lock != NULL)
2174 mutex_exit(hash_lock);
2176 /* Compute the hdr's checksum if necessary */
2177 arc_cksum_compute(buf);
2183 if (hdr_compressed == compressed) {
2184 if (!arc_buf_is_shared(buf)) {
2185 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2189 ASSERT(hdr_compressed);
2190 ASSERT(!compressed);
2191 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2194 * If the buf is sharing its data with the hdr, unlink it and
2195 * allocate a new data buffer for the buf.
2197 if (arc_buf_is_shared(buf)) {
2198 ASSERT(ARC_BUF_COMPRESSED(buf));
2200 /* We need to give the buf it's own b_data */
2201 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2203 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2204 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2206 /* Previously overhead was 0; just add new overhead */
2207 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2208 } else if (ARC_BUF_COMPRESSED(buf)) {
2209 /* We need to reallocate the buf's b_data */
2210 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2213 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2215 /* We increased the size of b_data; update overhead */
2216 ARCSTAT_INCR(arcstat_overhead_size,
2217 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2221 * Regardless of the buf's previous compression settings, it
2222 * should not be compressed at the end of this function.
2224 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2227 * Try copying the data from another buf which already has a
2228 * decompressed version. If that's not possible, it's time to
2229 * bite the bullet and decompress the data from the hdr.
2231 if (arc_buf_try_copy_decompressed_data(buf)) {
2232 /* Skip byteswapping and checksumming (already done) */
2233 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2236 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2237 hdr->b_l1hdr.b_pabd, buf->b_data,
2238 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2241 * Absent hardware errors or software bugs, this should
2242 * be impossible, but log it anyway so we can debug it.
2246 "hdr %p, compress %d, psize %d, lsize %d",
2247 hdr, arc_hdr_get_compress(hdr),
2248 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2249 if (hash_lock != NULL)
2250 mutex_enter(hash_lock);
2251 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2252 if (hash_lock != NULL)
2253 mutex_exit(hash_lock);
2254 return (SET_ERROR(EIO));
2260 /* Byteswap the buf's data if necessary */
2261 if (bswap != DMU_BSWAP_NUMFUNCS) {
2262 ASSERT(!HDR_SHARED_DATA(hdr));
2263 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2264 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2267 /* Compute the hdr's checksum if necessary */
2268 arc_cksum_compute(buf);
2274 * If this function is being called to decrypt an encrypted buffer or verify an
2275 * authenticated one, the key must be loaded and a mapping must be made
2276 * available in the keystore via spa_keystore_create_mapping() or one of its
2280 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2284 arc_fill_flags_t flags = 0;
2287 flags |= ARC_FILL_IN_PLACE;
2289 ret = arc_buf_fill(buf, spa, zb, flags);
2290 if (ret == ECKSUM) {
2292 * Convert authentication and decryption errors to EIO
2293 * (and generate an ereport) before leaving the ARC.
2295 ret = SET_ERROR(EIO);
2296 spa_log_error(spa, zb);
2297 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2298 spa, NULL, zb, NULL, 0, 0);
2305 * Increment the amount of evictable space in the arc_state_t's refcount.
2306 * We account for the space used by the hdr and the arc buf individually
2307 * so that we can add and remove them from the refcount individually.
2310 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2312 arc_buf_contents_t type = arc_buf_type(hdr);
2314 ASSERT(HDR_HAS_L1HDR(hdr));
2316 if (GHOST_STATE(state)) {
2317 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2318 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2319 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2320 ASSERT(!HDR_HAS_RABD(hdr));
2321 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2322 HDR_GET_LSIZE(hdr), hdr);
2326 ASSERT(!GHOST_STATE(state));
2327 if (hdr->b_l1hdr.b_pabd != NULL) {
2328 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2329 arc_hdr_size(hdr), hdr);
2331 if (HDR_HAS_RABD(hdr)) {
2332 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2333 HDR_GET_PSIZE(hdr), hdr);
2336 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2337 buf = buf->b_next) {
2338 if (arc_buf_is_shared(buf))
2340 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2341 arc_buf_size(buf), buf);
2346 * Decrement the amount of evictable space in the arc_state_t's refcount.
2347 * We account for the space used by the hdr and the arc buf individually
2348 * so that we can add and remove them from the refcount individually.
2351 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2353 arc_buf_contents_t type = arc_buf_type(hdr);
2355 ASSERT(HDR_HAS_L1HDR(hdr));
2357 if (GHOST_STATE(state)) {
2358 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2359 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2360 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2361 ASSERT(!HDR_HAS_RABD(hdr));
2362 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2363 HDR_GET_LSIZE(hdr), hdr);
2367 ASSERT(!GHOST_STATE(state));
2368 if (hdr->b_l1hdr.b_pabd != NULL) {
2369 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2370 arc_hdr_size(hdr), hdr);
2372 if (HDR_HAS_RABD(hdr)) {
2373 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2374 HDR_GET_PSIZE(hdr), hdr);
2377 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2378 buf = buf->b_next) {
2379 if (arc_buf_is_shared(buf))
2381 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2382 arc_buf_size(buf), buf);
2387 * Add a reference to this hdr indicating that someone is actively
2388 * referencing that memory. When the refcount transitions from 0 to 1,
2389 * we remove it from the respective arc_state_t list to indicate that
2390 * it is not evictable.
2393 add_reference(arc_buf_hdr_t *hdr, void *tag)
2397 ASSERT(HDR_HAS_L1HDR(hdr));
2398 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2399 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2400 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2401 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2404 state = hdr->b_l1hdr.b_state;
2406 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2407 (state != arc_anon)) {
2408 /* We don't use the L2-only state list. */
2409 if (state != arc_l2c_only) {
2410 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2412 arc_evictable_space_decrement(hdr, state);
2414 /* remove the prefetch flag if we get a reference */
2415 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2420 * Remove a reference from this hdr. When the reference transitions from
2421 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2422 * list making it eligible for eviction.
2425 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2428 arc_state_t *state = hdr->b_l1hdr.b_state;
2430 ASSERT(HDR_HAS_L1HDR(hdr));
2431 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2432 ASSERT(!GHOST_STATE(state));
2435 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2436 * check to prevent usage of the arc_l2c_only list.
2438 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2439 (state != arc_anon)) {
2440 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2441 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2442 arc_evictable_space_increment(hdr, state);
2448 * Returns detailed information about a specific arc buffer. When the
2449 * state_index argument is set the function will calculate the arc header
2450 * list position for its arc state. Since this requires a linear traversal
2451 * callers are strongly encourage not to do this. However, it can be helpful
2452 * for targeted analysis so the functionality is provided.
2455 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2457 arc_buf_hdr_t *hdr = ab->b_hdr;
2458 l1arc_buf_hdr_t *l1hdr = NULL;
2459 l2arc_buf_hdr_t *l2hdr = NULL;
2460 arc_state_t *state = NULL;
2462 memset(abi, 0, sizeof (arc_buf_info_t));
2467 abi->abi_flags = hdr->b_flags;
2469 if (HDR_HAS_L1HDR(hdr)) {
2470 l1hdr = &hdr->b_l1hdr;
2471 state = l1hdr->b_state;
2473 if (HDR_HAS_L2HDR(hdr))
2474 l2hdr = &hdr->b_l2hdr;
2477 abi->abi_bufcnt = l1hdr->b_bufcnt;
2478 abi->abi_access = l1hdr->b_arc_access;
2479 abi->abi_mru_hits = l1hdr->b_mru_hits;
2480 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2481 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2482 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2483 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2487 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2488 abi->abi_l2arc_hits = l2hdr->b_hits;
2491 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2492 abi->abi_state_contents = arc_buf_type(hdr);
2493 abi->abi_size = arc_hdr_size(hdr);
2497 * Move the supplied buffer to the indicated state. The hash lock
2498 * for the buffer must be held by the caller.
2501 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2502 kmutex_t *hash_lock)
2504 arc_state_t *old_state;
2507 boolean_t update_old, update_new;
2508 arc_buf_contents_t buftype = arc_buf_type(hdr);
2511 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2512 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2513 * L1 hdr doesn't always exist when we change state to arc_anon before
2514 * destroying a header, in which case reallocating to add the L1 hdr is
2517 if (HDR_HAS_L1HDR(hdr)) {
2518 old_state = hdr->b_l1hdr.b_state;
2519 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2520 bufcnt = hdr->b_l1hdr.b_bufcnt;
2521 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2524 old_state = arc_l2c_only;
2527 update_old = B_FALSE;
2529 update_new = update_old;
2531 ASSERT(MUTEX_HELD(hash_lock));
2532 ASSERT3P(new_state, !=, old_state);
2533 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2534 ASSERT(old_state != arc_anon || bufcnt <= 1);
2537 * If this buffer is evictable, transfer it from the
2538 * old state list to the new state list.
2541 if (old_state != arc_anon && old_state != arc_l2c_only) {
2542 ASSERT(HDR_HAS_L1HDR(hdr));
2543 multilist_remove(old_state->arcs_list[buftype], hdr);
2545 if (GHOST_STATE(old_state)) {
2547 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2548 update_old = B_TRUE;
2550 arc_evictable_space_decrement(hdr, old_state);
2552 if (new_state != arc_anon && new_state != arc_l2c_only) {
2554 * An L1 header always exists here, since if we're
2555 * moving to some L1-cached state (i.e. not l2c_only or
2556 * anonymous), we realloc the header to add an L1hdr
2559 ASSERT(HDR_HAS_L1HDR(hdr));
2560 multilist_insert(new_state->arcs_list[buftype], hdr);
2562 if (GHOST_STATE(new_state)) {
2564 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2565 update_new = B_TRUE;
2567 arc_evictable_space_increment(hdr, new_state);
2571 ASSERT(!HDR_EMPTY(hdr));
2572 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2573 buf_hash_remove(hdr);
2575 /* adjust state sizes (ignore arc_l2c_only) */
2577 if (update_new && new_state != arc_l2c_only) {
2578 ASSERT(HDR_HAS_L1HDR(hdr));
2579 if (GHOST_STATE(new_state)) {
2583 * When moving a header to a ghost state, we first
2584 * remove all arc buffers. Thus, we'll have a
2585 * bufcnt of zero, and no arc buffer to use for
2586 * the reference. As a result, we use the arc
2587 * header pointer for the reference.
2589 (void) zfs_refcount_add_many(&new_state->arcs_size,
2590 HDR_GET_LSIZE(hdr), hdr);
2591 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2592 ASSERT(!HDR_HAS_RABD(hdr));
2594 uint32_t buffers = 0;
2597 * Each individual buffer holds a unique reference,
2598 * thus we must remove each of these references one
2601 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2602 buf = buf->b_next) {
2603 ASSERT3U(bufcnt, !=, 0);
2607 * When the arc_buf_t is sharing the data
2608 * block with the hdr, the owner of the
2609 * reference belongs to the hdr. Only
2610 * add to the refcount if the arc_buf_t is
2613 if (arc_buf_is_shared(buf))
2616 (void) zfs_refcount_add_many(
2617 &new_state->arcs_size,
2618 arc_buf_size(buf), buf);
2620 ASSERT3U(bufcnt, ==, buffers);
2622 if (hdr->b_l1hdr.b_pabd != NULL) {
2623 (void) zfs_refcount_add_many(
2624 &new_state->arcs_size,
2625 arc_hdr_size(hdr), hdr);
2628 if (HDR_HAS_RABD(hdr)) {
2629 (void) zfs_refcount_add_many(
2630 &new_state->arcs_size,
2631 HDR_GET_PSIZE(hdr), hdr);
2636 if (update_old && old_state != arc_l2c_only) {
2637 ASSERT(HDR_HAS_L1HDR(hdr));
2638 if (GHOST_STATE(old_state)) {
2640 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2641 ASSERT(!HDR_HAS_RABD(hdr));
2644 * When moving a header off of a ghost state,
2645 * the header will not contain any arc buffers.
2646 * We use the arc header pointer for the reference
2647 * which is exactly what we did when we put the
2648 * header on the ghost state.
2651 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2652 HDR_GET_LSIZE(hdr), hdr);
2654 uint32_t buffers = 0;
2657 * Each individual buffer holds a unique reference,
2658 * thus we must remove each of these references one
2661 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2662 buf = buf->b_next) {
2663 ASSERT3U(bufcnt, !=, 0);
2667 * When the arc_buf_t is sharing the data
2668 * block with the hdr, the owner of the
2669 * reference belongs to the hdr. Only
2670 * add to the refcount if the arc_buf_t is
2673 if (arc_buf_is_shared(buf))
2676 (void) zfs_refcount_remove_many(
2677 &old_state->arcs_size, arc_buf_size(buf),
2680 ASSERT3U(bufcnt, ==, buffers);
2681 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2684 if (hdr->b_l1hdr.b_pabd != NULL) {
2685 (void) zfs_refcount_remove_many(
2686 &old_state->arcs_size, arc_hdr_size(hdr),
2690 if (HDR_HAS_RABD(hdr)) {
2691 (void) zfs_refcount_remove_many(
2692 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2698 if (HDR_HAS_L1HDR(hdr))
2699 hdr->b_l1hdr.b_state = new_state;
2702 * L2 headers should never be on the L2 state list since they don't
2703 * have L1 headers allocated.
2705 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2706 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2710 arc_space_consume(uint64_t space, arc_space_type_t type)
2712 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2717 case ARC_SPACE_DATA:
2718 aggsum_add(&astat_data_size, space);
2720 case ARC_SPACE_META:
2721 aggsum_add(&astat_metadata_size, space);
2723 case ARC_SPACE_BONUS:
2724 aggsum_add(&astat_bonus_size, space);
2726 case ARC_SPACE_DNODE:
2727 aggsum_add(&astat_dnode_size, space);
2729 case ARC_SPACE_DBUF:
2730 aggsum_add(&astat_dbuf_size, space);
2732 case ARC_SPACE_HDRS:
2733 aggsum_add(&astat_hdr_size, space);
2735 case ARC_SPACE_L2HDRS:
2736 aggsum_add(&astat_l2_hdr_size, space);
2740 if (type != ARC_SPACE_DATA)
2741 aggsum_add(&arc_meta_used, space);
2743 aggsum_add(&arc_size, space);
2747 arc_space_return(uint64_t space, arc_space_type_t type)
2749 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2754 case ARC_SPACE_DATA:
2755 aggsum_add(&astat_data_size, -space);
2757 case ARC_SPACE_META:
2758 aggsum_add(&astat_metadata_size, -space);
2760 case ARC_SPACE_BONUS:
2761 aggsum_add(&astat_bonus_size, -space);
2763 case ARC_SPACE_DNODE:
2764 aggsum_add(&astat_dnode_size, -space);
2766 case ARC_SPACE_DBUF:
2767 aggsum_add(&astat_dbuf_size, -space);
2769 case ARC_SPACE_HDRS:
2770 aggsum_add(&astat_hdr_size, -space);
2772 case ARC_SPACE_L2HDRS:
2773 aggsum_add(&astat_l2_hdr_size, -space);
2777 if (type != ARC_SPACE_DATA) {
2778 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2780 * We use the upper bound here rather than the precise value
2781 * because the arc_meta_max value doesn't need to be
2782 * precise. It's only consumed by humans via arcstats.
2784 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2785 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2786 aggsum_add(&arc_meta_used, -space);
2789 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2790 aggsum_add(&arc_size, -space);
2794 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2795 * with the hdr's b_pabd.
2798 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2801 * The criteria for sharing a hdr's data are:
2802 * 1. the buffer is not encrypted
2803 * 2. the hdr's compression matches the buf's compression
2804 * 3. the hdr doesn't need to be byteswapped
2805 * 4. the hdr isn't already being shared
2806 * 5. the buf is either compressed or it is the last buf in the hdr list
2808 * Criterion #5 maintains the invariant that shared uncompressed
2809 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2810 * might ask, "if a compressed buf is allocated first, won't that be the
2811 * last thing in the list?", but in that case it's impossible to create
2812 * a shared uncompressed buf anyway (because the hdr must be compressed
2813 * to have the compressed buf). You might also think that #3 is
2814 * sufficient to make this guarantee, however it's possible
2815 * (specifically in the rare L2ARC write race mentioned in
2816 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2817 * is sharable, but wasn't at the time of its allocation. Rather than
2818 * allow a new shared uncompressed buf to be created and then shuffle
2819 * the list around to make it the last element, this simply disallows
2820 * sharing if the new buf isn't the first to be added.
2822 ASSERT3P(buf->b_hdr, ==, hdr);
2823 boolean_t hdr_compressed =
2824 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2825 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2826 return (!ARC_BUF_ENCRYPTED(buf) &&
2827 buf_compressed == hdr_compressed &&
2828 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2829 !HDR_SHARED_DATA(hdr) &&
2830 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2834 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2835 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2836 * copy was made successfully, or an error code otherwise.
2839 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2840 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2841 boolean_t fill, arc_buf_t **ret)
2844 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2846 ASSERT(HDR_HAS_L1HDR(hdr));
2847 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2848 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2849 hdr->b_type == ARC_BUFC_METADATA);
2850 ASSERT3P(ret, !=, NULL);
2851 ASSERT3P(*ret, ==, NULL);
2852 IMPLY(encrypted, compressed);
2854 hdr->b_l1hdr.b_mru_hits = 0;
2855 hdr->b_l1hdr.b_mru_ghost_hits = 0;
2856 hdr->b_l1hdr.b_mfu_hits = 0;
2857 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
2858 hdr->b_l1hdr.b_l2_hits = 0;
2860 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2863 buf->b_next = hdr->b_l1hdr.b_buf;
2866 add_reference(hdr, tag);
2869 * We're about to change the hdr's b_flags. We must either
2870 * hold the hash_lock or be undiscoverable.
2872 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2875 * Only honor requests for compressed bufs if the hdr is actually
2876 * compressed. This must be overriden if the buffer is encrypted since
2877 * encrypted buffers cannot be decompressed.
2880 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2881 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2882 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2883 } else if (compressed &&
2884 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2885 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2886 flags |= ARC_FILL_COMPRESSED;
2891 flags |= ARC_FILL_NOAUTH;
2895 * If the hdr's data can be shared then we share the data buffer and
2896 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2897 * allocate a new buffer to store the buf's data.
2899 * There are two additional restrictions here because we're sharing
2900 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2901 * actively involved in an L2ARC write, because if this buf is used by
2902 * an arc_write() then the hdr's data buffer will be released when the
2903 * write completes, even though the L2ARC write might still be using it.
2904 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2905 * need to be ABD-aware.
2907 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2908 hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2910 /* Set up b_data and sharing */
2912 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2913 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2914 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2917 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2918 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2920 VERIFY3P(buf->b_data, !=, NULL);
2922 hdr->b_l1hdr.b_buf = buf;
2923 hdr->b_l1hdr.b_bufcnt += 1;
2925 hdr->b_crypt_hdr.b_ebufcnt += 1;
2928 * If the user wants the data from the hdr, we need to either copy or
2929 * decompress the data.
2932 ASSERT3P(zb, !=, NULL);
2933 return (arc_buf_fill(buf, spa, zb, flags));
2939 static char *arc_onloan_tag = "onloan";
2942 arc_loaned_bytes_update(int64_t delta)
2944 atomic_add_64(&arc_loaned_bytes, delta);
2946 /* assert that it did not wrap around */
2947 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2951 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2952 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2953 * buffers must be returned to the arc before they can be used by the DMU or
2957 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2959 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2960 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2962 arc_loaned_bytes_update(arc_buf_size(buf));
2968 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2969 enum zio_compress compression_type)
2971 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2972 psize, lsize, compression_type);
2974 arc_loaned_bytes_update(arc_buf_size(buf));
2980 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2981 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2982 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2983 enum zio_compress compression_type)
2985 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2986 byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
2988 atomic_add_64(&arc_loaned_bytes, psize);
2994 * Return a loaned arc buffer to the arc.
2997 arc_return_buf(arc_buf_t *buf, void *tag)
2999 arc_buf_hdr_t *hdr = buf->b_hdr;
3001 ASSERT3P(buf->b_data, !=, NULL);
3002 ASSERT(HDR_HAS_L1HDR(hdr));
3003 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
3004 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3006 arc_loaned_bytes_update(-arc_buf_size(buf));
3009 /* Detach an arc_buf from a dbuf (tag) */
3011 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3013 arc_buf_hdr_t *hdr = buf->b_hdr;
3015 ASSERT3P(buf->b_data, !=, NULL);
3016 ASSERT(HDR_HAS_L1HDR(hdr));
3017 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3018 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3020 arc_loaned_bytes_update(arc_buf_size(buf));
3024 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3026 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3029 df->l2df_size = size;
3030 df->l2df_type = type;
3031 mutex_enter(&l2arc_free_on_write_mtx);
3032 list_insert_head(l2arc_free_on_write, df);
3033 mutex_exit(&l2arc_free_on_write_mtx);
3037 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3039 arc_state_t *state = hdr->b_l1hdr.b_state;
3040 arc_buf_contents_t type = arc_buf_type(hdr);
3041 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3043 /* protected by hash lock, if in the hash table */
3044 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3045 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3046 ASSERT(state != arc_anon && state != arc_l2c_only);
3048 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
3051 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
3052 if (type == ARC_BUFC_METADATA) {
3053 arc_space_return(size, ARC_SPACE_META);
3055 ASSERT(type == ARC_BUFC_DATA);
3056 arc_space_return(size, ARC_SPACE_DATA);
3060 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
3062 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3067 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3068 * data buffer, we transfer the refcount ownership to the hdr and update
3069 * the appropriate kstats.
3072 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3074 ASSERT(arc_can_share(hdr, buf));
3075 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3076 ASSERT(!ARC_BUF_ENCRYPTED(buf));
3077 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3080 * Start sharing the data buffer. We transfer the
3081 * refcount ownership to the hdr since it always owns
3082 * the refcount whenever an arc_buf_t is shared.
3084 zfs_refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size,
3086 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3087 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3088 HDR_ISTYPE_METADATA(hdr));
3089 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3090 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3093 * Since we've transferred ownership to the hdr we need
3094 * to increment its compressed and uncompressed kstats and
3095 * decrement the overhead size.
3097 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3098 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3099 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3103 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3105 ASSERT(arc_buf_is_shared(buf));
3106 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3107 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3110 * We are no longer sharing this buffer so we need
3111 * to transfer its ownership to the rightful owner.
3113 zfs_refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size,
3115 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3116 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3117 abd_put(hdr->b_l1hdr.b_pabd);
3118 hdr->b_l1hdr.b_pabd = NULL;
3119 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3122 * Since the buffer is no longer shared between
3123 * the arc buf and the hdr, count it as overhead.
3125 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3126 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3127 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3131 * Remove an arc_buf_t from the hdr's buf list and return the last
3132 * arc_buf_t on the list. If no buffers remain on the list then return
3136 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3138 ASSERT(HDR_HAS_L1HDR(hdr));
3139 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3141 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3142 arc_buf_t *lastbuf = NULL;
3145 * Remove the buf from the hdr list and locate the last
3146 * remaining buffer on the list.
3148 while (*bufp != NULL) {
3150 *bufp = buf->b_next;
3153 * If we've removed a buffer in the middle of
3154 * the list then update the lastbuf and update
3157 if (*bufp != NULL) {
3159 bufp = &(*bufp)->b_next;
3163 ASSERT3P(lastbuf, !=, buf);
3164 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3165 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3166 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3172 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3176 arc_buf_destroy_impl(arc_buf_t *buf)
3178 arc_buf_hdr_t *hdr = buf->b_hdr;
3181 * Free up the data associated with the buf but only if we're not
3182 * sharing this with the hdr. If we are sharing it with the hdr, the
3183 * hdr is responsible for doing the free.
3185 if (buf->b_data != NULL) {
3187 * We're about to change the hdr's b_flags. We must either
3188 * hold the hash_lock or be undiscoverable.
3190 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3192 arc_cksum_verify(buf);
3193 arc_buf_unwatch(buf);
3195 if (arc_buf_is_shared(buf)) {
3196 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3198 uint64_t size = arc_buf_size(buf);
3199 arc_free_data_buf(hdr, buf->b_data, size, buf);
3200 ARCSTAT_INCR(arcstat_overhead_size, -size);
3204 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3205 hdr->b_l1hdr.b_bufcnt -= 1;
3207 if (ARC_BUF_ENCRYPTED(buf)) {
3208 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3211 * If we have no more encrypted buffers and we've
3212 * already gotten a copy of the decrypted data we can
3213 * free b_rabd to save some space.
3215 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3216 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3217 !HDR_IO_IN_PROGRESS(hdr)) {
3218 arc_hdr_free_abd(hdr, B_TRUE);
3223 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3225 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3227 * If the current arc_buf_t is sharing its data buffer with the
3228 * hdr, then reassign the hdr's b_pabd to share it with the new
3229 * buffer at the end of the list. The shared buffer is always
3230 * the last one on the hdr's buffer list.
3232 * There is an equivalent case for compressed bufs, but since
3233 * they aren't guaranteed to be the last buf in the list and
3234 * that is an exceedingly rare case, we just allow that space be
3235 * wasted temporarily. We must also be careful not to share
3236 * encrypted buffers, since they cannot be shared.
3238 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3239 /* Only one buf can be shared at once */
3240 VERIFY(!arc_buf_is_shared(lastbuf));
3241 /* hdr is uncompressed so can't have compressed buf */
3242 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3244 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3245 arc_hdr_free_abd(hdr, B_FALSE);
3248 * We must setup a new shared block between the
3249 * last buffer and the hdr. The data would have
3250 * been allocated by the arc buf so we need to transfer
3251 * ownership to the hdr since it's now being shared.
3253 arc_share_buf(hdr, lastbuf);
3255 } else if (HDR_SHARED_DATA(hdr)) {
3257 * Uncompressed shared buffers are always at the end
3258 * of the list. Compressed buffers don't have the
3259 * same requirements. This makes it hard to
3260 * simply assert that the lastbuf is shared so
3261 * we rely on the hdr's compression flags to determine
3262 * if we have a compressed, shared buffer.
3264 ASSERT3P(lastbuf, !=, NULL);
3265 ASSERT(arc_buf_is_shared(lastbuf) ||
3266 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3270 * Free the checksum if we're removing the last uncompressed buf from
3273 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3274 arc_cksum_free(hdr);
3277 /* clean up the buf */
3279 kmem_cache_free(buf_cache, buf);
3283 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata)
3287 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3288 ASSERT(HDR_HAS_L1HDR(hdr));
3289 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3290 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3293 size = HDR_GET_PSIZE(hdr);
3294 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3295 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr);
3296 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3297 ARCSTAT_INCR(arcstat_raw_size, size);
3299 size = arc_hdr_size(hdr);
3300 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3301 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr);
3302 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3305 ARCSTAT_INCR(arcstat_compressed_size, size);
3306 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3310 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3312 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3314 ASSERT(HDR_HAS_L1HDR(hdr));
3315 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3316 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3319 * If the hdr is currently being written to the l2arc then
3320 * we defer freeing the data by adding it to the l2arc_free_on_write
3321 * list. The l2arc will free the data once it's finished
3322 * writing it to the l2arc device.
3324 if (HDR_L2_WRITING(hdr)) {
3325 arc_hdr_free_on_write(hdr, free_rdata);
3326 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3327 } else if (free_rdata) {
3328 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3330 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3334 hdr->b_crypt_hdr.b_rabd = NULL;
3335 ARCSTAT_INCR(arcstat_raw_size, -size);
3337 hdr->b_l1hdr.b_pabd = NULL;
3340 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3341 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3343 ARCSTAT_INCR(arcstat_compressed_size, -size);
3344 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3347 static arc_buf_hdr_t *
3348 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3349 boolean_t protected, enum zio_compress compression_type,
3350 arc_buf_contents_t type, boolean_t alloc_rdata)
3354 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3356 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3358 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3361 ASSERT(HDR_EMPTY(hdr));
3362 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3363 HDR_SET_PSIZE(hdr, psize);
3364 HDR_SET_LSIZE(hdr, lsize);
3368 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3369 arc_hdr_set_compress(hdr, compression_type);
3371 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3373 hdr->b_l1hdr.b_state = arc_anon;
3374 hdr->b_l1hdr.b_arc_access = 0;
3375 hdr->b_l1hdr.b_bufcnt = 0;
3376 hdr->b_l1hdr.b_buf = NULL;
3379 * Allocate the hdr's buffer. This will contain either
3380 * the compressed or uncompressed data depending on the block
3381 * it references and compressed arc enablement.
3383 arc_hdr_alloc_abd(hdr, alloc_rdata);
3384 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3390 * Transition between the two allocation states for the arc_buf_hdr struct.
3391 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3392 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3393 * version is used when a cache buffer is only in the L2ARC in order to reduce
3396 static arc_buf_hdr_t *
3397 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3399 ASSERT(HDR_HAS_L2HDR(hdr));
3401 arc_buf_hdr_t *nhdr;
3402 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3404 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3405 (old == hdr_l2only_cache && new == hdr_full_cache));
3408 * if the caller wanted a new full header and the header is to be
3409 * encrypted we will actually allocate the header from the full crypt
3410 * cache instead. The same applies to freeing from the old cache.
3412 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3413 new = hdr_full_crypt_cache;
3414 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3415 old = hdr_full_crypt_cache;
3417 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3419 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3420 buf_hash_remove(hdr);
3422 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3424 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3425 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3427 * arc_access and arc_change_state need to be aware that a
3428 * header has just come out of L2ARC, so we set its state to
3429 * l2c_only even though it's about to change.
3431 nhdr->b_l1hdr.b_state = arc_l2c_only;
3433 /* Verify previous threads set to NULL before freeing */
3434 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3435 ASSERT(!HDR_HAS_RABD(hdr));
3437 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3438 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3439 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3442 * If we've reached here, We must have been called from
3443 * arc_evict_hdr(), as such we should have already been
3444 * removed from any ghost list we were previously on
3445 * (which protects us from racing with arc_evict_state),
3446 * thus no locking is needed during this check.
3448 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3451 * A buffer must not be moved into the arc_l2c_only
3452 * state if it's not finished being written out to the
3453 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3454 * might try to be accessed, even though it was removed.
3456 VERIFY(!HDR_L2_WRITING(hdr));
3457 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3458 ASSERT(!HDR_HAS_RABD(hdr));
3460 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3463 * The header has been reallocated so we need to re-insert it into any
3466 (void) buf_hash_insert(nhdr, NULL);
3468 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3470 mutex_enter(&dev->l2ad_mtx);
3473 * We must place the realloc'ed header back into the list at
3474 * the same spot. Otherwise, if it's placed earlier in the list,
3475 * l2arc_write_buffers() could find it during the function's
3476 * write phase, and try to write it out to the l2arc.
3478 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3479 list_remove(&dev->l2ad_buflist, hdr);
3481 mutex_exit(&dev->l2ad_mtx);
3484 * Since we're using the pointer address as the tag when
3485 * incrementing and decrementing the l2ad_alloc refcount, we
3486 * must remove the old pointer (that we're about to destroy) and
3487 * add the new pointer to the refcount. Otherwise we'd remove
3488 * the wrong pointer address when calling arc_hdr_destroy() later.
3491 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3492 arc_hdr_size(hdr), hdr);
3493 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3494 arc_hdr_size(nhdr), nhdr);
3496 buf_discard_identity(hdr);
3497 kmem_cache_free(old, hdr);
3503 * This function allows an L1 header to be reallocated as a crypt
3504 * header and vice versa. If we are going to a crypt header, the
3505 * new fields will be zeroed out.
3507 static arc_buf_hdr_t *
3508 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3510 arc_buf_hdr_t *nhdr;
3512 kmem_cache_t *ncache, *ocache;
3513 unsigned nsize, osize;
3516 * This function requires that hdr is in the arc_anon state.
3517 * Therefore it won't have any L2ARC data for us to worry
3520 ASSERT(HDR_HAS_L1HDR(hdr));
3521 ASSERT(!HDR_HAS_L2HDR(hdr));
3522 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3523 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3524 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3525 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3526 ASSERT3P(hdr->b_hash_next, ==, NULL);
3529 ncache = hdr_full_crypt_cache;
3530 nsize = sizeof (hdr->b_crypt_hdr);
3531 ocache = hdr_full_cache;
3532 osize = HDR_FULL_SIZE;
3534 ncache = hdr_full_cache;
3535 nsize = HDR_FULL_SIZE;
3536 ocache = hdr_full_crypt_cache;
3537 osize = sizeof (hdr->b_crypt_hdr);
3540 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3543 * Copy all members that aren't locks or condvars to the new header.
3544 * No lists are pointing to us (as we asserted above), so we don't
3545 * need to worry about the list nodes.
3547 nhdr->b_dva = hdr->b_dva;
3548 nhdr->b_birth = hdr->b_birth;
3549 nhdr->b_type = hdr->b_type;
3550 nhdr->b_flags = hdr->b_flags;
3551 nhdr->b_psize = hdr->b_psize;
3552 nhdr->b_lsize = hdr->b_lsize;
3553 nhdr->b_spa = hdr->b_spa;
3554 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3555 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3556 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3557 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3558 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3559 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3560 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3561 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3562 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3563 nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits;
3564 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3565 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3568 * This zfs_refcount_add() exists only to ensure that the individual
3569 * arc buffers always point to a header that is referenced, avoiding
3570 * a small race condition that could trigger ASSERTs.
3572 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3573 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3574 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3575 mutex_enter(&buf->b_evict_lock);
3577 mutex_exit(&buf->b_evict_lock);
3580 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3581 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3582 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3585 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3587 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3590 /* unset all members of the original hdr */
3591 bzero(&hdr->b_dva, sizeof (dva_t));
3593 hdr->b_type = ARC_BUFC_INVALID;
3598 hdr->b_l1hdr.b_freeze_cksum = NULL;
3599 hdr->b_l1hdr.b_buf = NULL;
3600 hdr->b_l1hdr.b_bufcnt = 0;
3601 hdr->b_l1hdr.b_byteswap = 0;
3602 hdr->b_l1hdr.b_state = NULL;
3603 hdr->b_l1hdr.b_arc_access = 0;
3604 hdr->b_l1hdr.b_mru_hits = 0;
3605 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3606 hdr->b_l1hdr.b_mfu_hits = 0;
3607 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3608 hdr->b_l1hdr.b_l2_hits = 0;
3609 hdr->b_l1hdr.b_acb = NULL;
3610 hdr->b_l1hdr.b_pabd = NULL;
3612 if (ocache == hdr_full_crypt_cache) {
3613 ASSERT(!HDR_HAS_RABD(hdr));
3614 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3615 hdr->b_crypt_hdr.b_ebufcnt = 0;
3616 hdr->b_crypt_hdr.b_dsobj = 0;
3617 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3618 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3619 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3622 buf_discard_identity(hdr);
3623 kmem_cache_free(ocache, hdr);
3629 * This function is used by the send / receive code to convert a newly
3630 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3631 * is also used to allow the root objset block to be uupdated without altering
3632 * its embedded MACs. Both block types will always be uncompressed so we do not
3633 * have to worry about compression type or psize.
3636 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3637 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3640 arc_buf_hdr_t *hdr = buf->b_hdr;
3642 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3643 ASSERT(HDR_HAS_L1HDR(hdr));
3644 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3646 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3647 if (!HDR_PROTECTED(hdr))
3648 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3649 hdr->b_crypt_hdr.b_dsobj = dsobj;
3650 hdr->b_crypt_hdr.b_ot = ot;
3651 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3652 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3653 if (!arc_hdr_has_uncompressed_buf(hdr))
3654 arc_cksum_free(hdr);
3657 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3659 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3661 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3665 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3666 * The buf is returned thawed since we expect the consumer to modify it.
3669 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3671 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3672 B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3673 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3675 arc_buf_t *buf = NULL;
3676 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3677 B_FALSE, B_FALSE, &buf));
3684 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3685 * for bufs containing metadata.
3688 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3689 enum zio_compress compression_type)
3691 ASSERT3U(lsize, >, 0);
3692 ASSERT3U(lsize, >=, psize);
3693 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3694 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3696 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3697 B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3698 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3700 arc_buf_t *buf = NULL;
3701 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3702 B_TRUE, B_FALSE, B_FALSE, &buf));
3704 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3706 if (!arc_buf_is_shared(buf)) {
3708 * To ensure that the hdr has the correct data in it if we call
3709 * arc_untransform() on this buf before it's been written to
3710 * disk, it's easiest if we just set up sharing between the
3713 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3714 arc_hdr_free_abd(hdr, B_FALSE);
3715 arc_share_buf(hdr, buf);
3722 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3723 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3724 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3725 enum zio_compress compression_type)
3729 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3730 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3732 ASSERT3U(lsize, >, 0);
3733 ASSERT3U(lsize, >=, psize);
3734 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3735 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3737 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3738 compression_type, type, B_TRUE);
3739 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3741 hdr->b_crypt_hdr.b_dsobj = dsobj;
3742 hdr->b_crypt_hdr.b_ot = ot;
3743 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3744 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3745 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3746 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3747 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3750 * This buffer will be considered encrypted even if the ot is not an
3751 * encrypted type. It will become authenticated instead in
3752 * arc_write_ready().
3755 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3756 B_FALSE, B_FALSE, &buf));
3758 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3764 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3766 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3767 l2arc_dev_t *dev = l2hdr->b_dev;
3768 uint64_t psize = arc_hdr_size(hdr);
3770 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3771 ASSERT(HDR_HAS_L2HDR(hdr));
3773 list_remove(&dev->l2ad_buflist, hdr);
3775 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3776 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3778 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3780 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3781 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3785 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3787 if (HDR_HAS_L1HDR(hdr)) {
3788 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3789 hdr->b_l1hdr.b_bufcnt > 0);
3790 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3791 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3793 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3794 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3796 if (!HDR_EMPTY(hdr))
3797 buf_discard_identity(hdr);
3799 if (HDR_HAS_L2HDR(hdr)) {
3800 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3801 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3804 mutex_enter(&dev->l2ad_mtx);
3807 * Even though we checked this conditional above, we
3808 * need to check this again now that we have the
3809 * l2ad_mtx. This is because we could be racing with
3810 * another thread calling l2arc_evict() which might have
3811 * destroyed this header's L2 portion as we were waiting
3812 * to acquire the l2ad_mtx. If that happens, we don't
3813 * want to re-destroy the header's L2 portion.
3815 if (HDR_HAS_L2HDR(hdr))
3816 arc_hdr_l2hdr_destroy(hdr);
3819 mutex_exit(&dev->l2ad_mtx);
3822 if (HDR_HAS_L1HDR(hdr)) {
3823 arc_cksum_free(hdr);
3825 while (hdr->b_l1hdr.b_buf != NULL)
3826 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3828 if (hdr->b_l1hdr.b_pabd != NULL) {
3829 arc_hdr_free_abd(hdr, B_FALSE);
3832 if (HDR_HAS_RABD(hdr))
3833 arc_hdr_free_abd(hdr, B_TRUE);
3836 ASSERT3P(hdr->b_hash_next, ==, NULL);
3837 if (HDR_HAS_L1HDR(hdr)) {
3838 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3839 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3841 if (!HDR_PROTECTED(hdr)) {
3842 kmem_cache_free(hdr_full_cache, hdr);
3844 kmem_cache_free(hdr_full_crypt_cache, hdr);
3847 kmem_cache_free(hdr_l2only_cache, hdr);
3852 arc_buf_destroy(arc_buf_t *buf, void* tag)
3854 arc_buf_hdr_t *hdr = buf->b_hdr;
3855 kmutex_t *hash_lock = HDR_LOCK(hdr);
3857 if (hdr->b_l1hdr.b_state == arc_anon) {
3858 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3859 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3860 VERIFY0(remove_reference(hdr, NULL, tag));
3861 arc_hdr_destroy(hdr);
3865 mutex_enter(hash_lock);
3866 ASSERT3P(hdr, ==, buf->b_hdr);
3867 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3868 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3869 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3870 ASSERT3P(buf->b_data, !=, NULL);
3872 (void) remove_reference(hdr, hash_lock, tag);
3873 arc_buf_destroy_impl(buf);
3874 mutex_exit(hash_lock);
3878 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3879 * state of the header is dependent on its state prior to entering this
3880 * function. The following transitions are possible:
3882 * - arc_mru -> arc_mru_ghost
3883 * - arc_mfu -> arc_mfu_ghost
3884 * - arc_mru_ghost -> arc_l2c_only
3885 * - arc_mru_ghost -> deleted
3886 * - arc_mfu_ghost -> arc_l2c_only
3887 * - arc_mfu_ghost -> deleted
3890 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3892 arc_state_t *evicted_state, *state;
3893 int64_t bytes_evicted = 0;
3894 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3895 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3897 ASSERT(MUTEX_HELD(hash_lock));
3898 ASSERT(HDR_HAS_L1HDR(hdr));
3900 state = hdr->b_l1hdr.b_state;
3901 if (GHOST_STATE(state)) {
3902 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3903 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3906 * l2arc_write_buffers() relies on a header's L1 portion
3907 * (i.e. its b_pabd field) during it's write phase.
3908 * Thus, we cannot push a header onto the arc_l2c_only
3909 * state (removing its L1 piece) until the header is
3910 * done being written to the l2arc.
3912 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3913 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3914 return (bytes_evicted);
3917 ARCSTAT_BUMP(arcstat_deleted);
3918 bytes_evicted += HDR_GET_LSIZE(hdr);
3920 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3922 if (HDR_HAS_L2HDR(hdr)) {
3923 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3924 ASSERT(!HDR_HAS_RABD(hdr));
3926 * This buffer is cached on the 2nd Level ARC;
3927 * don't destroy the header.
3929 arc_change_state(arc_l2c_only, hdr, hash_lock);
3931 * dropping from L1+L2 cached to L2-only,
3932 * realloc to remove the L1 header.
3934 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3937 arc_change_state(arc_anon, hdr, hash_lock);
3938 arc_hdr_destroy(hdr);
3940 return (bytes_evicted);
3943 ASSERT(state == arc_mru || state == arc_mfu);
3944 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3946 /* prefetch buffers have a minimum lifespan */
3947 if (HDR_IO_IN_PROGRESS(hdr) ||
3948 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3949 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3950 MSEC_TO_TICK(min_lifetime))) {
3951 ARCSTAT_BUMP(arcstat_evict_skip);
3952 return (bytes_evicted);
3955 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3956 while (hdr->b_l1hdr.b_buf) {
3957 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3958 if (!mutex_tryenter(&buf->b_evict_lock)) {
3959 ARCSTAT_BUMP(arcstat_mutex_miss);
3962 if (buf->b_data != NULL)
3963 bytes_evicted += HDR_GET_LSIZE(hdr);
3964 mutex_exit(&buf->b_evict_lock);
3965 arc_buf_destroy_impl(buf);
3968 if (HDR_HAS_L2HDR(hdr)) {
3969 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3971 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3972 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3973 HDR_GET_LSIZE(hdr));
3975 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3976 HDR_GET_LSIZE(hdr));
3980 if (hdr->b_l1hdr.b_bufcnt == 0) {
3981 arc_cksum_free(hdr);
3983 bytes_evicted += arc_hdr_size(hdr);
3986 * If this hdr is being evicted and has a compressed
3987 * buffer then we discard it here before we change states.
3988 * This ensures that the accounting is updated correctly
3989 * in arc_free_data_impl().
3991 if (hdr->b_l1hdr.b_pabd != NULL)
3992 arc_hdr_free_abd(hdr, B_FALSE);
3994 if (HDR_HAS_RABD(hdr))
3995 arc_hdr_free_abd(hdr, B_TRUE);
3997 arc_change_state(evicted_state, hdr, hash_lock);
3998 ASSERT(HDR_IN_HASH_TABLE(hdr));
3999 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
4000 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4003 return (bytes_evicted);
4007 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4008 uint64_t spa, int64_t bytes)
4010 multilist_sublist_t *mls;
4011 uint64_t bytes_evicted = 0;
4013 kmutex_t *hash_lock;
4014 int evict_count = 0;
4016 ASSERT3P(marker, !=, NULL);
4017 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4019 mls = multilist_sublist_lock(ml, idx);
4021 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
4022 hdr = multilist_sublist_prev(mls, marker)) {
4023 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
4024 (evict_count >= zfs_arc_evict_batch_limit))
4028 * To keep our iteration location, move the marker
4029 * forward. Since we're not holding hdr's hash lock, we
4030 * must be very careful and not remove 'hdr' from the
4031 * sublist. Otherwise, other consumers might mistake the
4032 * 'hdr' as not being on a sublist when they call the
4033 * multilist_link_active() function (they all rely on
4034 * the hash lock protecting concurrent insertions and
4035 * removals). multilist_sublist_move_forward() was
4036 * specifically implemented to ensure this is the case
4037 * (only 'marker' will be removed and re-inserted).
4039 multilist_sublist_move_forward(mls, marker);
4042 * The only case where the b_spa field should ever be
4043 * zero, is the marker headers inserted by
4044 * arc_evict_state(). It's possible for multiple threads
4045 * to be calling arc_evict_state() concurrently (e.g.
4046 * dsl_pool_close() and zio_inject_fault()), so we must
4047 * skip any markers we see from these other threads.
4049 if (hdr->b_spa == 0)
4052 /* we're only interested in evicting buffers of a certain spa */
4053 if (spa != 0 && hdr->b_spa != spa) {
4054 ARCSTAT_BUMP(arcstat_evict_skip);
4058 hash_lock = HDR_LOCK(hdr);
4061 * We aren't calling this function from any code path
4062 * that would already be holding a hash lock, so we're
4063 * asserting on this assumption to be defensive in case
4064 * this ever changes. Without this check, it would be
4065 * possible to incorrectly increment arcstat_mutex_miss
4066 * below (e.g. if the code changed such that we called
4067 * this function with a hash lock held).
4069 ASSERT(!MUTEX_HELD(hash_lock));
4071 if (mutex_tryenter(hash_lock)) {
4072 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
4073 mutex_exit(hash_lock);
4075 bytes_evicted += evicted;
4078 * If evicted is zero, arc_evict_hdr() must have
4079 * decided to skip this header, don't increment
4080 * evict_count in this case.
4086 * If arc_size isn't overflowing, signal any
4087 * threads that might happen to be waiting.
4089 * For each header evicted, we wake up a single
4090 * thread. If we used cv_broadcast, we could
4091 * wake up "too many" threads causing arc_size
4092 * to significantly overflow arc_c; since
4093 * arc_get_data_impl() doesn't check for overflow
4094 * when it's woken up (it doesn't because it's
4095 * possible for the ARC to be overflowing while
4096 * full of un-evictable buffers, and the
4097 * function should proceed in this case).
4099 * If threads are left sleeping, due to not
4100 * using cv_broadcast, they will be woken up
4101 * just before arc_reclaim_thread() sleeps.
4103 mutex_enter(&arc_reclaim_lock);
4104 if (!arc_is_overflowing())
4105 cv_signal(&arc_reclaim_waiters_cv);
4106 mutex_exit(&arc_reclaim_lock);
4108 ARCSTAT_BUMP(arcstat_mutex_miss);
4112 multilist_sublist_unlock(mls);
4114 return (bytes_evicted);
4118 * Evict buffers from the given arc state, until we've removed the
4119 * specified number of bytes. Move the removed buffers to the
4120 * appropriate evict state.
4122 * This function makes a "best effort". It skips over any buffers
4123 * it can't get a hash_lock on, and so, may not catch all candidates.
4124 * It may also return without evicting as much space as requested.
4126 * If bytes is specified using the special value ARC_EVICT_ALL, this
4127 * will evict all available (i.e. unlocked and evictable) buffers from
4128 * the given arc state; which is used by arc_flush().
4131 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4132 arc_buf_contents_t type)
4134 uint64_t total_evicted = 0;
4135 multilist_t *ml = state->arcs_list[type];
4137 arc_buf_hdr_t **markers;
4139 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4141 num_sublists = multilist_get_num_sublists(ml);
4144 * If we've tried to evict from each sublist, made some
4145 * progress, but still have not hit the target number of bytes
4146 * to evict, we want to keep trying. The markers allow us to
4147 * pick up where we left off for each individual sublist, rather
4148 * than starting from the tail each time.
4150 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4151 for (int i = 0; i < num_sublists; i++) {
4152 multilist_sublist_t *mls;
4154 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4157 * A b_spa of 0 is used to indicate that this header is
4158 * a marker. This fact is used in arc_adjust_type() and
4159 * arc_evict_state_impl().
4161 markers[i]->b_spa = 0;
4163 mls = multilist_sublist_lock(ml, i);
4164 multilist_sublist_insert_tail(mls, markers[i]);
4165 multilist_sublist_unlock(mls);
4169 * While we haven't hit our target number of bytes to evict, or
4170 * we're evicting all available buffers.
4172 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4173 int sublist_idx = multilist_get_random_index(ml);
4174 uint64_t scan_evicted = 0;
4177 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4178 * Request that 10% of the LRUs be scanned by the superblock
4181 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
4182 arc_dnode_limit) > 0) {
4183 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
4184 arc_dnode_limit) / sizeof (dnode_t) /
4185 zfs_arc_dnode_reduce_percent);
4189 * Start eviction using a randomly selected sublist,
4190 * this is to try and evenly balance eviction across all
4191 * sublists. Always starting at the same sublist
4192 * (e.g. index 0) would cause evictions to favor certain
4193 * sublists over others.
4195 for (int i = 0; i < num_sublists; i++) {
4196 uint64_t bytes_remaining;
4197 uint64_t bytes_evicted;
4199 if (bytes == ARC_EVICT_ALL)
4200 bytes_remaining = ARC_EVICT_ALL;
4201 else if (total_evicted < bytes)
4202 bytes_remaining = bytes - total_evicted;
4206 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4207 markers[sublist_idx], spa, bytes_remaining);
4209 scan_evicted += bytes_evicted;
4210 total_evicted += bytes_evicted;
4212 /* we've reached the end, wrap to the beginning */
4213 if (++sublist_idx >= num_sublists)
4218 * If we didn't evict anything during this scan, we have
4219 * no reason to believe we'll evict more during another
4220 * scan, so break the loop.
4222 if (scan_evicted == 0) {
4223 /* This isn't possible, let's make that obvious */
4224 ASSERT3S(bytes, !=, 0);
4227 * When bytes is ARC_EVICT_ALL, the only way to
4228 * break the loop is when scan_evicted is zero.
4229 * In that case, we actually have evicted enough,
4230 * so we don't want to increment the kstat.
4232 if (bytes != ARC_EVICT_ALL) {
4233 ASSERT3S(total_evicted, <, bytes);
4234 ARCSTAT_BUMP(arcstat_evict_not_enough);
4241 for (int i = 0; i < num_sublists; i++) {
4242 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4243 multilist_sublist_remove(mls, markers[i]);
4244 multilist_sublist_unlock(mls);
4246 kmem_cache_free(hdr_full_cache, markers[i]);
4248 kmem_free(markers, sizeof (*markers) * num_sublists);
4250 return (total_evicted);
4254 * Flush all "evictable" data of the given type from the arc state
4255 * specified. This will not evict any "active" buffers (i.e. referenced).
4257 * When 'retry' is set to B_FALSE, the function will make a single pass
4258 * over the state and evict any buffers that it can. Since it doesn't
4259 * continually retry the eviction, it might end up leaving some buffers
4260 * in the ARC due to lock misses.
4262 * When 'retry' is set to B_TRUE, the function will continually retry the
4263 * eviction until *all* evictable buffers have been removed from the
4264 * state. As a result, if concurrent insertions into the state are
4265 * allowed (e.g. if the ARC isn't shutting down), this function might
4266 * wind up in an infinite loop, continually trying to evict buffers.
4269 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4272 uint64_t evicted = 0;
4274 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4275 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4285 * Helper function for arc_prune_async() it is responsible for safely
4286 * handling the execution of a registered arc_prune_func_t.
4289 arc_prune_task(void *ptr)
4291 arc_prune_t *ap = (arc_prune_t *)ptr;
4292 arc_prune_func_t *func = ap->p_pfunc;
4295 func(ap->p_adjust, ap->p_private);
4297 zfs_refcount_remove(&ap->p_refcnt, func);
4301 * Notify registered consumers they must drop holds on a portion of the ARC
4302 * buffered they reference. This provides a mechanism to ensure the ARC can
4303 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4304 * is analogous to dnlc_reduce_cache() but more generic.
4306 * This operation is performed asynchronously so it may be safely called
4307 * in the context of the arc_reclaim_thread(). A reference is taken here
4308 * for each registered arc_prune_t and the arc_prune_task() is responsible
4309 * for releasing it once the registered arc_prune_func_t has completed.
4312 arc_prune_async(int64_t adjust)
4316 mutex_enter(&arc_prune_mtx);
4317 for (ap = list_head(&arc_prune_list); ap != NULL;
4318 ap = list_next(&arc_prune_list, ap)) {
4320 if (zfs_refcount_count(&ap->p_refcnt) >= 2)
4323 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
4324 ap->p_adjust = adjust;
4325 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4326 ap, TQ_SLEEP) == TASKQID_INVALID) {
4327 zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4330 ARCSTAT_BUMP(arcstat_prune);
4332 mutex_exit(&arc_prune_mtx);
4336 * Evict the specified number of bytes from the state specified,
4337 * restricting eviction to the spa and type given. This function
4338 * prevents us from trying to evict more from a state's list than
4339 * is "evictable", and to skip evicting altogether when passed a
4340 * negative value for "bytes". In contrast, arc_evict_state() will
4341 * evict everything it can, when passed a negative value for "bytes".
4344 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4345 arc_buf_contents_t type)
4349 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4350 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4352 return (arc_evict_state(state, spa, delta, type));
4359 * The goal of this function is to evict enough meta data buffers from the
4360 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4361 * more complicated than it appears because it is common for data buffers
4362 * to have holds on meta data buffers. In addition, dnode meta data buffers
4363 * will be held by the dnodes in the block preventing them from being freed.
4364 * This means we can't simply traverse the ARC and expect to always find
4365 * enough unheld meta data buffer to release.
4367 * Therefore, this function has been updated to make alternating passes
4368 * over the ARC releasing data buffers and then newly unheld meta data
4369 * buffers. This ensures forward progress is maintained and meta_used
4370 * will decrease. Normally this is sufficient, but if required the ARC
4371 * will call the registered prune callbacks causing dentry and inodes to
4372 * be dropped from the VFS cache. This will make dnode meta data buffers
4373 * available for reclaim.
4376 arc_adjust_meta_balanced(uint64_t meta_used)
4378 int64_t delta, prune = 0, adjustmnt;
4379 uint64_t total_evicted = 0;
4380 arc_buf_contents_t type = ARC_BUFC_DATA;
4381 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4385 * This slightly differs than the way we evict from the mru in
4386 * arc_adjust because we don't have a "target" value (i.e. no
4387 * "meta" arc_p). As a result, I think we can completely
4388 * cannibalize the metadata in the MRU before we evict the
4389 * metadata from the MFU. I think we probably need to implement a
4390 * "metadata arc_p" value to do this properly.
4392 adjustmnt = meta_used - arc_meta_limit;
4394 if (adjustmnt > 0 &&
4395 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4396 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4398 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4403 * We can't afford to recalculate adjustmnt here. If we do,
4404 * new metadata buffers can sneak into the MRU or ANON lists,
4405 * thus penalize the MFU metadata. Although the fudge factor is
4406 * small, it has been empirically shown to be significant for
4407 * certain workloads (e.g. creating many empty directories). As
4408 * such, we use the original calculation for adjustmnt, and
4409 * simply decrement the amount of data evicted from the MRU.
4412 if (adjustmnt > 0 &&
4413 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4414 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4416 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4419 adjustmnt = meta_used - arc_meta_limit;
4421 if (adjustmnt > 0 &&
4422 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4423 delta = MIN(adjustmnt,
4424 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4425 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4429 if (adjustmnt > 0 &&
4430 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4431 delta = MIN(adjustmnt,
4432 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4433 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4437 * If after attempting to make the requested adjustment to the ARC
4438 * the meta limit is still being exceeded then request that the
4439 * higher layers drop some cached objects which have holds on ARC
4440 * meta buffers. Requests to the upper layers will be made with
4441 * increasingly large scan sizes until the ARC is below the limit.
4443 if (meta_used > arc_meta_limit) {
4444 if (type == ARC_BUFC_DATA) {
4445 type = ARC_BUFC_METADATA;
4447 type = ARC_BUFC_DATA;
4449 if (zfs_arc_meta_prune) {
4450 prune += zfs_arc_meta_prune;
4451 arc_prune_async(prune);
4460 return (total_evicted);
4464 * Evict metadata buffers from the cache, such that arc_meta_used is
4465 * capped by the arc_meta_limit tunable.
4468 arc_adjust_meta_only(uint64_t meta_used)
4470 uint64_t total_evicted = 0;
4474 * If we're over the meta limit, we want to evict enough
4475 * metadata to get back under the meta limit. We don't want to
4476 * evict so much that we drop the MRU below arc_p, though. If
4477 * we're over the meta limit more than we're over arc_p, we
4478 * evict some from the MRU here, and some from the MFU below.
4480 target = MIN((int64_t)(meta_used - arc_meta_limit),
4481 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4482 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4484 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4487 * Similar to the above, we want to evict enough bytes to get us
4488 * below the meta limit, but not so much as to drop us below the
4489 * space allotted to the MFU (which is defined as arc_c - arc_p).
4491 target = MIN((int64_t)(meta_used - arc_meta_limit),
4492 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4495 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4497 return (total_evicted);
4501 arc_adjust_meta(uint64_t meta_used)
4503 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4504 return (arc_adjust_meta_only(meta_used));
4506 return (arc_adjust_meta_balanced(meta_used));
4510 * Return the type of the oldest buffer in the given arc state
4512 * This function will select a random sublist of type ARC_BUFC_DATA and
4513 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4514 * is compared, and the type which contains the "older" buffer will be
4517 static arc_buf_contents_t
4518 arc_adjust_type(arc_state_t *state)
4520 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4521 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4522 int data_idx = multilist_get_random_index(data_ml);
4523 int meta_idx = multilist_get_random_index(meta_ml);
4524 multilist_sublist_t *data_mls;
4525 multilist_sublist_t *meta_mls;
4526 arc_buf_contents_t type;
4527 arc_buf_hdr_t *data_hdr;
4528 arc_buf_hdr_t *meta_hdr;
4531 * We keep the sublist lock until we're finished, to prevent
4532 * the headers from being destroyed via arc_evict_state().
4534 data_mls = multilist_sublist_lock(data_ml, data_idx);
4535 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4538 * These two loops are to ensure we skip any markers that
4539 * might be at the tail of the lists due to arc_evict_state().
4542 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4543 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4544 if (data_hdr->b_spa != 0)
4548 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4549 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4550 if (meta_hdr->b_spa != 0)
4554 if (data_hdr == NULL && meta_hdr == NULL) {
4555 type = ARC_BUFC_DATA;
4556 } else if (data_hdr == NULL) {
4557 ASSERT3P(meta_hdr, !=, NULL);
4558 type = ARC_BUFC_METADATA;
4559 } else if (meta_hdr == NULL) {
4560 ASSERT3P(data_hdr, !=, NULL);
4561 type = ARC_BUFC_DATA;
4563 ASSERT3P(data_hdr, !=, NULL);
4564 ASSERT3P(meta_hdr, !=, NULL);
4566 /* The headers can't be on the sublist without an L1 header */
4567 ASSERT(HDR_HAS_L1HDR(data_hdr));
4568 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4570 if (data_hdr->b_l1hdr.b_arc_access <
4571 meta_hdr->b_l1hdr.b_arc_access) {
4572 type = ARC_BUFC_DATA;
4574 type = ARC_BUFC_METADATA;
4578 multilist_sublist_unlock(meta_mls);
4579 multilist_sublist_unlock(data_mls);
4585 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4590 uint64_t total_evicted = 0;
4593 uint64_t asize = aggsum_value(&arc_size);
4594 uint64_t ameta = aggsum_value(&arc_meta_used);
4597 * If we're over arc_meta_limit, we want to correct that before
4598 * potentially evicting data buffers below.
4600 total_evicted += arc_adjust_meta(ameta);
4605 * If we're over the target cache size, we want to evict enough
4606 * from the list to get back to our target size. We don't want
4607 * to evict too much from the MRU, such that it drops below
4608 * arc_p. So, if we're over our target cache size more than
4609 * the MRU is over arc_p, we'll evict enough to get back to
4610 * arc_p here, and then evict more from the MFU below.
4612 target = MIN((int64_t)(asize - arc_c),
4613 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4614 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4617 * If we're below arc_meta_min, always prefer to evict data.
4618 * Otherwise, try to satisfy the requested number of bytes to
4619 * evict from the type which contains older buffers; in an
4620 * effort to keep newer buffers in the cache regardless of their
4621 * type. If we cannot satisfy the number of bytes from this
4622 * type, spill over into the next type.
4624 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4625 ameta > arc_meta_min) {
4626 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4627 total_evicted += bytes;
4630 * If we couldn't evict our target number of bytes from
4631 * metadata, we try to get the rest from data.
4636 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4638 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4639 total_evicted += bytes;
4642 * If we couldn't evict our target number of bytes from
4643 * data, we try to get the rest from metadata.
4648 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4652 * Re-sum ARC stats after the first round of evictions.
4654 asize = aggsum_value(&arc_size);
4655 ameta = aggsum_value(&arc_meta_used);
4661 * Now that we've tried to evict enough from the MRU to get its
4662 * size back to arc_p, if we're still above the target cache
4663 * size, we evict the rest from the MFU.
4665 target = asize - arc_c;
4667 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4668 ameta > arc_meta_min) {
4669 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4670 total_evicted += bytes;
4673 * If we couldn't evict our target number of bytes from
4674 * metadata, we try to get the rest from data.
4679 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4681 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4682 total_evicted += bytes;
4685 * If we couldn't evict our target number of bytes from
4686 * data, we try to get the rest from data.
4691 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4695 * Adjust ghost lists
4697 * In addition to the above, the ARC also defines target values
4698 * for the ghost lists. The sum of the mru list and mru ghost
4699 * list should never exceed the target size of the cache, and
4700 * the sum of the mru list, mfu list, mru ghost list, and mfu
4701 * ghost list should never exceed twice the target size of the
4702 * cache. The following logic enforces these limits on the ghost
4703 * caches, and evicts from them as needed.
4705 target = zfs_refcount_count(&arc_mru->arcs_size) +
4706 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4708 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4709 total_evicted += bytes;
4714 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4717 * We assume the sum of the mru list and mfu list is less than
4718 * or equal to arc_c (we enforced this above), which means we
4719 * can use the simpler of the two equations below:
4721 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4722 * mru ghost + mfu ghost <= arc_c
4724 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4725 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4727 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4728 total_evicted += bytes;
4733 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4735 return (total_evicted);
4739 arc_flush(spa_t *spa, boolean_t retry)
4744 * If retry is B_TRUE, a spa must not be specified since we have
4745 * no good way to determine if all of a spa's buffers have been
4746 * evicted from an arc state.
4748 ASSERT(!retry || spa == 0);
4751 guid = spa_load_guid(spa);
4753 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4754 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4756 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4757 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4759 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4760 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4762 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4763 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4767 arc_shrink(int64_t to_free)
4769 uint64_t asize = aggsum_value(&arc_size);
4772 if (c > to_free && c - to_free > arc_c_min) {
4773 arc_c = c - to_free;
4774 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4776 arc_c = MAX(asize, arc_c_min);
4778 arc_p = (arc_c >> 1);
4779 ASSERT(arc_c >= arc_c_min);
4780 ASSERT((int64_t)arc_p >= 0);
4786 (void) arc_adjust();
4790 * Return maximum amount of memory that we could possibly use. Reduced
4791 * to half of all memory in user space which is primarily used for testing.
4794 arc_all_memory(void)
4797 #ifdef CONFIG_HIGHMEM
4798 return (ptob(totalram_pages - totalhigh_pages));
4800 return (ptob(totalram_pages));
4801 #endif /* CONFIG_HIGHMEM */
4803 return (ptob(physmem) / 2);
4804 #endif /* _KERNEL */
4808 * Return the amount of memory that is considered free. In user space
4809 * which is primarily used for testing we pretend that free memory ranges
4810 * from 0-20% of all memory.
4813 arc_free_memory(void)
4816 #ifdef CONFIG_HIGHMEM
4819 return (ptob(si.freeram - si.freehigh));
4821 return (ptob(nr_free_pages() +
4822 nr_inactive_file_pages() +
4823 nr_inactive_anon_pages() +
4824 nr_slab_reclaimable_pages()));
4826 #endif /* CONFIG_HIGHMEM */
4828 return (spa_get_random(arc_all_memory() * 20 / 100));
4829 #endif /* _KERNEL */
4832 typedef enum free_memory_reason_t {
4837 FMR_PAGES_PP_MAXIMUM,
4840 } free_memory_reason_t;
4842 int64_t last_free_memory;
4843 free_memory_reason_t last_free_reason;
4847 * Additional reserve of pages for pp_reserve.
4849 int64_t arc_pages_pp_reserve = 64;
4852 * Additional reserve of pages for swapfs.
4854 int64_t arc_swapfs_reserve = 64;
4855 #endif /* _KERNEL */
4858 * Return the amount of memory that can be consumed before reclaim will be
4859 * needed. Positive if there is sufficient free memory, negative indicates
4860 * the amount of memory that needs to be freed up.
4863 arc_available_memory(void)
4865 int64_t lowest = INT64_MAX;
4866 free_memory_reason_t r = FMR_UNKNOWN;
4873 pgcnt_t needfree = btop(arc_need_free);
4874 pgcnt_t lotsfree = btop(arc_sys_free);
4875 pgcnt_t desfree = 0;
4876 pgcnt_t freemem = btop(arc_free_memory());
4880 n = PAGESIZE * (-needfree);
4888 * check that we're out of range of the pageout scanner. It starts to
4889 * schedule paging if freemem is less than lotsfree and needfree.
4890 * lotsfree is the high-water mark for pageout, and needfree is the
4891 * number of needed free pages. We add extra pages here to make sure
4892 * the scanner doesn't start up while we're freeing memory.
4894 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4902 * check to make sure that swapfs has enough space so that anon
4903 * reservations can still succeed. anon_resvmem() checks that the
4904 * availrmem is greater than swapfs_minfree, and the number of reserved
4905 * swap pages. We also add a bit of extra here just to prevent
4906 * circumstances from getting really dire.
4908 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4909 desfree - arc_swapfs_reserve);
4912 r = FMR_SWAPFS_MINFREE;
4916 * Check that we have enough availrmem that memory locking (e.g., via
4917 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4918 * stores the number of pages that cannot be locked; when availrmem
4919 * drops below pages_pp_maximum, page locking mechanisms such as
4920 * page_pp_lock() will fail.)
4922 n = PAGESIZE * (availrmem - pages_pp_maximum -
4923 arc_pages_pp_reserve);
4926 r = FMR_PAGES_PP_MAXIMUM;
4932 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4933 * kernel heap space before we ever run out of available physical
4934 * memory. Most checks of the size of the heap_area compare against
4935 * tune.t_minarmem, which is the minimum available real memory that we
4936 * can have in the system. However, this is generally fixed at 25 pages
4937 * which is so low that it's useless. In this comparison, we seek to
4938 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4939 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4942 n = vmem_size(heap_arena, VMEM_FREE) -
4943 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4951 * If zio data pages are being allocated out of a separate heap segment,
4952 * then enforce that the size of available vmem for this arena remains
4953 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4955 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4956 * memory (in the zio_arena) free, which can avoid memory
4957 * fragmentation issues.
4959 if (zio_arena != NULL) {
4960 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4961 (vmem_size(zio_arena, VMEM_ALLOC) >>
4962 arc_zio_arena_free_shift);
4969 /* Every 100 calls, free a small amount */
4970 if (spa_get_random(100) == 0)
4972 #endif /* _KERNEL */
4974 last_free_memory = lowest;
4975 last_free_reason = r;
4981 * Determine if the system is under memory pressure and is asking
4982 * to reclaim memory. A return value of B_TRUE indicates that the system
4983 * is under memory pressure and that the arc should adjust accordingly.
4986 arc_reclaim_needed(void)
4988 return (arc_available_memory() < 0);
4992 arc_kmem_reap_now(void)
4995 kmem_cache_t *prev_cache = NULL;
4996 kmem_cache_t *prev_data_cache = NULL;
4997 extern kmem_cache_t *zio_buf_cache[];
4998 extern kmem_cache_t *zio_data_buf_cache[];
4999 extern kmem_cache_t *range_seg_cache;
5002 if ((aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) &&
5003 zfs_arc_meta_prune) {
5005 * We are exceeding our meta-data cache limit.
5006 * Prune some entries to release holds on meta-data.
5008 arc_prune_async(zfs_arc_meta_prune);
5012 * Reclaim unused memory from all kmem caches.
5018 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
5020 /* reach upper limit of cache size on 32-bit */
5021 if (zio_buf_cache[i] == NULL)
5024 if (zio_buf_cache[i] != prev_cache) {
5025 prev_cache = zio_buf_cache[i];
5026 kmem_cache_reap_now(zio_buf_cache[i]);
5028 if (zio_data_buf_cache[i] != prev_data_cache) {
5029 prev_data_cache = zio_data_buf_cache[i];
5030 kmem_cache_reap_now(zio_data_buf_cache[i]);
5033 kmem_cache_reap_now(buf_cache);
5034 kmem_cache_reap_now(hdr_full_cache);
5035 kmem_cache_reap_now(hdr_l2only_cache);
5036 kmem_cache_reap_now(range_seg_cache);
5038 if (zio_arena != NULL) {
5040 * Ask the vmem arena to reclaim unused memory from its
5043 vmem_qcache_reap(zio_arena);
5048 * Threads can block in arc_get_data_impl() waiting for this thread to evict
5049 * enough data and signal them to proceed. When this happens, the threads in
5050 * arc_get_data_impl() are sleeping while holding the hash lock for their
5051 * particular arc header. Thus, we must be careful to never sleep on a
5052 * hash lock in this thread. This is to prevent the following deadlock:
5054 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
5055 * waiting for the reclaim thread to signal it.
5057 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
5058 * fails, and goes to sleep forever.
5060 * This possible deadlock is avoided by always acquiring a hash lock
5061 * using mutex_tryenter() from arc_reclaim_thread().
5065 arc_reclaim_thread(void *unused)
5067 fstrans_cookie_t cookie = spl_fstrans_mark();
5068 hrtime_t growtime = 0;
5071 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
5073 mutex_enter(&arc_reclaim_lock);
5074 while (!arc_reclaim_thread_exit) {
5075 uint64_t evicted = 0;
5076 uint64_t need_free = arc_need_free;
5077 arc_tuning_update();
5080 * This is necessary in order for the mdb ::arc dcmd to
5081 * show up to date information. Since the ::arc command
5082 * does not call the kstat's update function, without
5083 * this call, the command may show stale stats for the
5084 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
5085 * with this change, the data might be up to 1 second
5086 * out of date; but that should suffice. The arc_state_t
5087 * structures can be queried directly if more accurate
5088 * information is needed.
5091 if (arc_ksp != NULL)
5092 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
5094 mutex_exit(&arc_reclaim_lock);
5097 * We call arc_adjust() before (possibly) calling
5098 * arc_kmem_reap_now(), so that we can wake up
5099 * arc_get_data_buf() sooner.
5101 evicted = arc_adjust();
5103 int64_t free_memory = arc_available_memory();
5104 if (free_memory < 0) {
5106 arc_no_grow = B_TRUE;
5110 * Wait at least zfs_grow_retry (default 5) seconds
5111 * before considering growing.
5113 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
5115 arc_kmem_reap_now();
5118 * If we are still low on memory, shrink the ARC
5119 * so that we have arc_shrink_min free space.
5121 free_memory = arc_available_memory();
5124 (arc_c >> arc_shrink_shift) - free_memory;
5127 to_free = MAX(to_free, need_free);
5129 arc_shrink(to_free);
5131 } else if (free_memory < arc_c >> arc_no_grow_shift) {
5132 arc_no_grow = B_TRUE;
5133 } else if (gethrtime() >= growtime) {
5134 arc_no_grow = B_FALSE;
5137 mutex_enter(&arc_reclaim_lock);
5140 * If evicted is zero, we couldn't evict anything via
5141 * arc_adjust(). This could be due to hash lock
5142 * collisions, but more likely due to the majority of
5143 * arc buffers being unevictable. Therefore, even if
5144 * arc_size is above arc_c, another pass is unlikely to
5145 * be helpful and could potentially cause us to enter an
5148 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
5150 * We're either no longer overflowing, or we
5151 * can't evict anything more, so we should wake
5152 * up any threads before we go to sleep and remove
5153 * the bytes we were working on from arc_need_free
5154 * since nothing more will be done here.
5156 cv_broadcast(&arc_reclaim_waiters_cv);
5157 ARCSTAT_INCR(arcstat_need_free, -need_free);
5160 * Block until signaled, or after one second (we
5161 * might need to perform arc_kmem_reap_now()
5162 * even if we aren't being signalled)
5164 CALLB_CPR_SAFE_BEGIN(&cpr);
5165 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv,
5166 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
5167 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
5171 arc_reclaim_thread_exit = B_FALSE;
5172 cv_broadcast(&arc_reclaim_thread_cv);
5173 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
5174 spl_fstrans_unmark(cookie);
5180 * Determine the amount of memory eligible for eviction contained in the
5181 * ARC. All clean data reported by the ghost lists can always be safely
5182 * evicted. Due to arc_c_min, the same does not hold for all clean data
5183 * contained by the regular mru and mfu lists.
5185 * In the case of the regular mru and mfu lists, we need to report as
5186 * much clean data as possible, such that evicting that same reported
5187 * data will not bring arc_size below arc_c_min. Thus, in certain
5188 * circumstances, the total amount of clean data in the mru and mfu
5189 * lists might not actually be evictable.
5191 * The following two distinct cases are accounted for:
5193 * 1. The sum of the amount of dirty data contained by both the mru and
5194 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5195 * is greater than or equal to arc_c_min.
5196 * (i.e. amount of dirty data >= arc_c_min)
5198 * This is the easy case; all clean data contained by the mru and mfu
5199 * lists is evictable. Evicting all clean data can only drop arc_size
5200 * to the amount of dirty data, which is greater than arc_c_min.
5202 * 2. The sum of the amount of dirty data contained by both the mru and
5203 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5204 * is less than arc_c_min.
5205 * (i.e. arc_c_min > amount of dirty data)
5207 * 2.1. arc_size is greater than or equal arc_c_min.
5208 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5210 * In this case, not all clean data from the regular mru and mfu
5211 * lists is actually evictable; we must leave enough clean data
5212 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5213 * evictable data from the two lists combined, is exactly the
5214 * difference between arc_size and arc_c_min.
5216 * 2.2. arc_size is less than arc_c_min
5217 * (i.e. arc_c_min > arc_size > amount of dirty data)
5219 * In this case, none of the data contained in the mru and mfu
5220 * lists is evictable, even if it's clean. Since arc_size is
5221 * already below arc_c_min, evicting any more would only
5222 * increase this negative difference.
5225 arc_evictable_memory(void)
5227 int64_t asize = aggsum_value(&arc_size);
5228 uint64_t arc_clean =
5229 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
5230 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
5231 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
5232 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5233 uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0);
5236 * Scale reported evictable memory in proportion to page cache, cap
5237 * at specified min/max.
5239 uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent;
5240 min = MAX(arc_c_min, MIN(arc_c_max, min));
5242 if (arc_dirty >= min)
5245 return (MAX((int64_t)asize - (int64_t)min, 0));
5249 * If sc->nr_to_scan is zero, the caller is requesting a query of the
5250 * number of objects which can potentially be freed. If it is nonzero,
5251 * the request is to free that many objects.
5253 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
5254 * in struct shrinker and also require the shrinker to return the number
5257 * Older kernels require the shrinker to return the number of freeable
5258 * objects following the freeing of nr_to_free.
5260 static spl_shrinker_t
5261 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
5265 /* The arc is considered warm once reclaim has occurred */
5266 if (unlikely(arc_warm == B_FALSE))
5269 /* Return the potential number of reclaimable pages */
5270 pages = btop((int64_t)arc_evictable_memory());
5271 if (sc->nr_to_scan == 0)
5274 /* Not allowed to perform filesystem reclaim */
5275 if (!(sc->gfp_mask & __GFP_FS))
5276 return (SHRINK_STOP);
5278 /* Reclaim in progress */
5279 if (mutex_tryenter(&arc_reclaim_lock) == 0) {
5280 ARCSTAT_INCR(arcstat_need_free, ptob(sc->nr_to_scan));
5284 mutex_exit(&arc_reclaim_lock);
5287 * Evict the requested number of pages by shrinking arc_c the
5291 arc_shrink(ptob(sc->nr_to_scan));
5292 if (current_is_kswapd())
5293 arc_kmem_reap_now();
5294 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
5295 pages = MAX((int64_t)pages -
5296 (int64_t)btop(arc_evictable_memory()), 0);
5298 pages = btop(arc_evictable_memory());
5301 * We've shrunk what we can, wake up threads.
5303 cv_broadcast(&arc_reclaim_waiters_cv);
5305 pages = SHRINK_STOP;
5308 * When direct reclaim is observed it usually indicates a rapid
5309 * increase in memory pressure. This occurs because the kswapd
5310 * threads were unable to asynchronously keep enough free memory
5311 * available. In this case set arc_no_grow to briefly pause arc
5312 * growth to avoid compounding the memory pressure.
5314 if (current_is_kswapd()) {
5315 ARCSTAT_BUMP(arcstat_memory_indirect_count);
5317 arc_no_grow = B_TRUE;
5318 arc_kmem_reap_now();
5319 ARCSTAT_BUMP(arcstat_memory_direct_count);
5324 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
5326 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
5327 #endif /* _KERNEL */
5330 * Adapt arc info given the number of bytes we are trying to add and
5331 * the state that we are coming from. This function is only called
5332 * when we are adding new content to the cache.
5335 arc_adapt(int bytes, arc_state_t *state)
5338 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5339 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5340 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5342 if (state == arc_l2c_only)
5347 * Adapt the target size of the MRU list:
5348 * - if we just hit in the MRU ghost list, then increase
5349 * the target size of the MRU list.
5350 * - if we just hit in the MFU ghost list, then increase
5351 * the target size of the MFU list by decreasing the
5352 * target size of the MRU list.
5354 if (state == arc_mru_ghost) {
5355 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5356 if (!zfs_arc_p_dampener_disable)
5357 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5359 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5360 } else if (state == arc_mfu_ghost) {
5363 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5364 if (!zfs_arc_p_dampener_disable)
5365 mult = MIN(mult, 10);
5367 delta = MIN(bytes * mult, arc_p);
5368 arc_p = MAX(arc_p_min, arc_p - delta);
5370 ASSERT((int64_t)arc_p >= 0);
5372 if (arc_reclaim_needed()) {
5373 cv_signal(&arc_reclaim_thread_cv);
5380 if (arc_c >= arc_c_max)
5384 * If we're within (2 * maxblocksize) bytes of the target
5385 * cache size, increment the target cache size
5387 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5388 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >=
5390 atomic_add_64(&arc_c, (int64_t)bytes);
5391 if (arc_c > arc_c_max)
5393 else if (state == arc_anon)
5394 atomic_add_64(&arc_p, (int64_t)bytes);
5398 ASSERT((int64_t)arc_p >= 0);
5402 * Check if arc_size has grown past our upper threshold, determined by
5403 * zfs_arc_overflow_shift.
5406 arc_is_overflowing(void)
5408 /* Always allow at least one block of overflow */
5409 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5410 arc_c >> zfs_arc_overflow_shift);
5413 * We just compare the lower bound here for performance reasons. Our
5414 * primary goals are to make sure that the arc never grows without
5415 * bound, and that it can reach its maximum size. This check
5416 * accomplishes both goals. The maximum amount we could run over by is
5417 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5418 * in the ARC. In practice, that's in the tens of MB, which is low
5419 * enough to be safe.
5421 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5425 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5427 arc_buf_contents_t type = arc_buf_type(hdr);
5429 arc_get_data_impl(hdr, size, tag);
5430 if (type == ARC_BUFC_METADATA) {
5431 return (abd_alloc(size, B_TRUE));
5433 ASSERT(type == ARC_BUFC_DATA);
5434 return (abd_alloc(size, B_FALSE));
5439 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5441 arc_buf_contents_t type = arc_buf_type(hdr);
5443 arc_get_data_impl(hdr, size, tag);
5444 if (type == ARC_BUFC_METADATA) {
5445 return (zio_buf_alloc(size));
5447 ASSERT(type == ARC_BUFC_DATA);
5448 return (zio_data_buf_alloc(size));
5453 * Allocate a block and return it to the caller. If we are hitting the
5454 * hard limit for the cache size, we must sleep, waiting for the eviction
5455 * thread to catch up. If we're past the target size but below the hard
5456 * limit, we'll only signal the reclaim thread and continue on.
5459 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5461 arc_state_t *state = hdr->b_l1hdr.b_state;
5462 arc_buf_contents_t type = arc_buf_type(hdr);
5464 arc_adapt(size, state);
5467 * If arc_size is currently overflowing, and has grown past our
5468 * upper limit, we must be adding data faster than the evict
5469 * thread can evict. Thus, to ensure we don't compound the
5470 * problem by adding more data and forcing arc_size to grow even
5471 * further past it's target size, we halt and wait for the
5472 * eviction thread to catch up.
5474 * It's also possible that the reclaim thread is unable to evict
5475 * enough buffers to get arc_size below the overflow limit (e.g.
5476 * due to buffers being un-evictable, or hash lock collisions).
5477 * In this case, we want to proceed regardless if we're
5478 * overflowing; thus we don't use a while loop here.
5480 if (arc_is_overflowing()) {
5481 mutex_enter(&arc_reclaim_lock);
5484 * Now that we've acquired the lock, we may no longer be
5485 * over the overflow limit, lets check.
5487 * We're ignoring the case of spurious wake ups. If that
5488 * were to happen, it'd let this thread consume an ARC
5489 * buffer before it should have (i.e. before we're under
5490 * the overflow limit and were signalled by the reclaim
5491 * thread). As long as that is a rare occurrence, it
5492 * shouldn't cause any harm.
5494 if (arc_is_overflowing()) {
5495 cv_signal(&arc_reclaim_thread_cv);
5496 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5499 mutex_exit(&arc_reclaim_lock);
5502 VERIFY3U(hdr->b_type, ==, type);
5503 if (type == ARC_BUFC_METADATA) {
5504 arc_space_consume(size, ARC_SPACE_META);
5506 arc_space_consume(size, ARC_SPACE_DATA);
5510 * Update the state size. Note that ghost states have a
5511 * "ghost size" and so don't need to be updated.
5513 if (!GHOST_STATE(state)) {
5515 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5518 * If this is reached via arc_read, the link is
5519 * protected by the hash lock. If reached via
5520 * arc_buf_alloc, the header should not be accessed by
5521 * any other thread. And, if reached via arc_read_done,
5522 * the hash lock will protect it if it's found in the
5523 * hash table; otherwise no other thread should be
5524 * trying to [add|remove]_reference it.
5526 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5527 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5528 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5533 * If we are growing the cache, and we are adding anonymous
5534 * data, and we have outgrown arc_p, update arc_p
5536 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5537 hdr->b_l1hdr.b_state == arc_anon &&
5538 (zfs_refcount_count(&arc_anon->arcs_size) +
5539 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5540 arc_p = MIN(arc_c, arc_p + size);
5545 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5547 arc_free_data_impl(hdr, size, tag);
5552 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5554 arc_buf_contents_t type = arc_buf_type(hdr);
5556 arc_free_data_impl(hdr, size, tag);
5557 if (type == ARC_BUFC_METADATA) {
5558 zio_buf_free(buf, size);
5560 ASSERT(type == ARC_BUFC_DATA);
5561 zio_data_buf_free(buf, size);
5566 * Free the arc data buffer.
5569 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5571 arc_state_t *state = hdr->b_l1hdr.b_state;
5572 arc_buf_contents_t type = arc_buf_type(hdr);
5574 /* protected by hash lock, if in the hash table */
5575 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5576 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5577 ASSERT(state != arc_anon && state != arc_l2c_only);
5579 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5582 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5584 VERIFY3U(hdr->b_type, ==, type);
5585 if (type == ARC_BUFC_METADATA) {
5586 arc_space_return(size, ARC_SPACE_META);
5588 ASSERT(type == ARC_BUFC_DATA);
5589 arc_space_return(size, ARC_SPACE_DATA);
5594 * This routine is called whenever a buffer is accessed.
5595 * NOTE: the hash lock is dropped in this function.
5598 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5602 ASSERT(MUTEX_HELD(hash_lock));
5603 ASSERT(HDR_HAS_L1HDR(hdr));
5605 if (hdr->b_l1hdr.b_state == arc_anon) {
5607 * This buffer is not in the cache, and does not
5608 * appear in our "ghost" list. Add the new buffer
5612 ASSERT0(hdr->b_l1hdr.b_arc_access);
5613 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5614 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5615 arc_change_state(arc_mru, hdr, hash_lock);
5617 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5618 now = ddi_get_lbolt();
5621 * If this buffer is here because of a prefetch, then either:
5622 * - clear the flag if this is a "referencing" read
5623 * (any subsequent access will bump this into the MFU state).
5625 * - move the buffer to the head of the list if this is
5626 * another prefetch (to make it less likely to be evicted).
5628 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5629 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5630 /* link protected by hash lock */
5631 ASSERT(multilist_link_active(
5632 &hdr->b_l1hdr.b_arc_node));
5634 arc_hdr_clear_flags(hdr,
5636 ARC_FLAG_PRESCIENT_PREFETCH);
5637 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5638 ARCSTAT_BUMP(arcstat_mru_hits);
5640 hdr->b_l1hdr.b_arc_access = now;
5645 * This buffer has been "accessed" only once so far,
5646 * but it is still in the cache. Move it to the MFU
5649 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5652 * More than 125ms have passed since we
5653 * instantiated this buffer. Move it to the
5654 * most frequently used state.
5656 hdr->b_l1hdr.b_arc_access = now;
5657 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5658 arc_change_state(arc_mfu, hdr, hash_lock);
5660 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5661 ARCSTAT_BUMP(arcstat_mru_hits);
5662 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5663 arc_state_t *new_state;
5665 * This buffer has been "accessed" recently, but
5666 * was evicted from the cache. Move it to the
5670 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5671 new_state = arc_mru;
5672 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5673 arc_hdr_clear_flags(hdr,
5675 ARC_FLAG_PRESCIENT_PREFETCH);
5677 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5679 new_state = arc_mfu;
5680 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5683 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5684 arc_change_state(new_state, hdr, hash_lock);
5686 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5687 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5688 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5690 * This buffer has been accessed more than once and is
5691 * still in the cache. Keep it in the MFU state.
5693 * NOTE: an add_reference() that occurred when we did
5694 * the arc_read() will have kicked this off the list.
5695 * If it was a prefetch, we will explicitly move it to
5696 * the head of the list now.
5699 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5700 ARCSTAT_BUMP(arcstat_mfu_hits);
5701 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5702 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5703 arc_state_t *new_state = arc_mfu;
5705 * This buffer has been accessed more than once but has
5706 * been evicted from the cache. Move it back to the
5710 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5712 * This is a prefetch access...
5713 * move this block back to the MRU state.
5715 new_state = arc_mru;
5718 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5719 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5720 arc_change_state(new_state, hdr, hash_lock);
5722 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5723 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5724 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5726 * This buffer is on the 2nd Level ARC.
5729 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5730 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5731 arc_change_state(arc_mfu, hdr, hash_lock);
5733 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5734 hdr->b_l1hdr.b_state);
5739 * This routine is called by dbuf_hold() to update the arc_access() state
5740 * which otherwise would be skipped for entries in the dbuf cache.
5743 arc_buf_access(arc_buf_t *buf)
5745 mutex_enter(&buf->b_evict_lock);
5746 arc_buf_hdr_t *hdr = buf->b_hdr;
5749 * Avoid taking the hash_lock when possible as an optimization.
5750 * The header must be checked again under the hash_lock in order
5751 * to handle the case where it is concurrently being released.
5753 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5754 mutex_exit(&buf->b_evict_lock);
5758 kmutex_t *hash_lock = HDR_LOCK(hdr);
5759 mutex_enter(hash_lock);
5761 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5762 mutex_exit(hash_lock);
5763 mutex_exit(&buf->b_evict_lock);
5764 ARCSTAT_BUMP(arcstat_access_skip);
5768 mutex_exit(&buf->b_evict_lock);
5770 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5771 hdr->b_l1hdr.b_state == arc_mfu);
5773 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5774 arc_access(hdr, hash_lock);
5775 mutex_exit(hash_lock);
5777 ARCSTAT_BUMP(arcstat_hits);
5778 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5779 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5782 /* a generic arc_read_done_func_t which you can use */
5785 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5786 arc_buf_t *buf, void *arg)
5791 bcopy(buf->b_data, arg, arc_buf_size(buf));
5792 arc_buf_destroy(buf, arg);
5795 /* a generic arc_read_done_func_t */
5798 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5799 arc_buf_t *buf, void *arg)
5801 arc_buf_t **bufp = arg;
5804 ASSERT(zio == NULL || zio->io_error != 0);
5807 ASSERT(zio == NULL || zio->io_error == 0);
5809 ASSERT(buf->b_data != NULL);
5814 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5816 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5817 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5818 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5820 if (HDR_COMPRESSION_ENABLED(hdr)) {
5821 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5822 BP_GET_COMPRESS(bp));
5824 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5825 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5826 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5831 arc_read_done(zio_t *zio)
5833 blkptr_t *bp = zio->io_bp;
5834 arc_buf_hdr_t *hdr = zio->io_private;
5835 kmutex_t *hash_lock = NULL;
5836 arc_callback_t *callback_list;
5837 arc_callback_t *acb;
5838 boolean_t freeable = B_FALSE;
5841 * The hdr was inserted into hash-table and removed from lists
5842 * prior to starting I/O. We should find this header, since
5843 * it's in the hash table, and it should be legit since it's
5844 * not possible to evict it during the I/O. The only possible
5845 * reason for it not to be found is if we were freed during the
5848 if (HDR_IN_HASH_TABLE(hdr)) {
5849 arc_buf_hdr_t *found;
5851 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5852 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5853 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5854 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5855 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5857 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5859 ASSERT((found == hdr &&
5860 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5861 (found == hdr && HDR_L2_READING(hdr)));
5862 ASSERT3P(hash_lock, !=, NULL);
5865 if (BP_IS_PROTECTED(bp)) {
5866 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5867 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5868 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5869 hdr->b_crypt_hdr.b_iv);
5871 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5874 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5875 sizeof (zil_chain_t));
5876 zio_crypt_decode_mac_zil(tmpbuf,
5877 hdr->b_crypt_hdr.b_mac);
5878 abd_return_buf(zio->io_abd, tmpbuf,
5879 sizeof (zil_chain_t));
5881 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5885 if (zio->io_error == 0) {
5886 /* byteswap if necessary */
5887 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5888 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5889 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5891 hdr->b_l1hdr.b_byteswap =
5892 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5895 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5899 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5900 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5901 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5903 callback_list = hdr->b_l1hdr.b_acb;
5904 ASSERT3P(callback_list, !=, NULL);
5906 if (hash_lock && zio->io_error == 0 &&
5907 hdr->b_l1hdr.b_state == arc_anon) {
5909 * Only call arc_access on anonymous buffers. This is because
5910 * if we've issued an I/O for an evicted buffer, we've already
5911 * called arc_access (to prevent any simultaneous readers from
5912 * getting confused).
5914 arc_access(hdr, hash_lock);
5918 * If a read request has a callback (i.e. acb_done is not NULL), then we
5919 * make a buf containing the data according to the parameters which were
5920 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5921 * aren't needlessly decompressing the data multiple times.
5923 int callback_cnt = 0;
5924 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5930 if (zio->io_error != 0)
5933 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5934 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5935 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5939 * Assert non-speculative zios didn't fail because an
5940 * encryption key wasn't loaded
5942 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5946 * If we failed to decrypt, report an error now (as the zio
5947 * layer would have done if it had done the transforms).
5949 if (error == ECKSUM) {
5950 ASSERT(BP_IS_PROTECTED(bp));
5951 error = SET_ERROR(EIO);
5952 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5953 spa_log_error(zio->io_spa, &acb->acb_zb);
5954 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
5955 zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
5961 * Decompression or decryption failed. Set
5962 * io_error so that when we call acb_done
5963 * (below), we will indicate that the read
5964 * failed. Note that in the unusual case
5965 * where one callback is compressed and another
5966 * uncompressed, we will mark all of them
5967 * as failed, even though the uncompressed
5968 * one can't actually fail. In this case,
5969 * the hdr will not be anonymous, because
5970 * if there are multiple callbacks, it's
5971 * because multiple threads found the same
5972 * arc buf in the hash table.
5974 zio->io_error = error;
5979 * If there are multiple callbacks, we must have the hash lock,
5980 * because the only way for multiple threads to find this hdr is
5981 * in the hash table. This ensures that if there are multiple
5982 * callbacks, the hdr is not anonymous. If it were anonymous,
5983 * we couldn't use arc_buf_destroy() in the error case below.
5985 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5987 hdr->b_l1hdr.b_acb = NULL;
5988 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5989 if (callback_cnt == 0)
5990 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5992 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5993 callback_list != NULL);
5995 if (zio->io_error == 0) {
5996 arc_hdr_verify(hdr, zio->io_bp);
5998 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5999 if (hdr->b_l1hdr.b_state != arc_anon)
6000 arc_change_state(arc_anon, hdr, hash_lock);
6001 if (HDR_IN_HASH_TABLE(hdr))
6002 buf_hash_remove(hdr);
6003 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
6007 * Broadcast before we drop the hash_lock to avoid the possibility
6008 * that the hdr (and hence the cv) might be freed before we get to
6009 * the cv_broadcast().
6011 cv_broadcast(&hdr->b_l1hdr.b_cv);
6013 if (hash_lock != NULL) {
6014 mutex_exit(hash_lock);
6017 * This block was freed while we waited for the read to
6018 * complete. It has been removed from the hash table and
6019 * moved to the anonymous state (so that it won't show up
6022 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
6023 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
6026 /* execute each callback and free its structure */
6027 while ((acb = callback_list) != NULL) {
6028 if (acb->acb_done != NULL) {
6029 if (zio->io_error != 0 && acb->acb_buf != NULL) {
6031 * If arc_buf_alloc_impl() fails during
6032 * decompression, the buf will still be
6033 * allocated, and needs to be freed here.
6035 arc_buf_destroy(acb->acb_buf,
6037 acb->acb_buf = NULL;
6039 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
6040 acb->acb_buf, acb->acb_private);
6043 if (acb->acb_zio_dummy != NULL) {
6044 acb->acb_zio_dummy->io_error = zio->io_error;
6045 zio_nowait(acb->acb_zio_dummy);
6048 callback_list = acb->acb_next;
6049 kmem_free(acb, sizeof (arc_callback_t));
6053 arc_hdr_destroy(hdr);
6057 * "Read" the block at the specified DVA (in bp) via the
6058 * cache. If the block is found in the cache, invoke the provided
6059 * callback immediately and return. Note that the `zio' parameter
6060 * in the callback will be NULL in this case, since no IO was
6061 * required. If the block is not in the cache pass the read request
6062 * on to the spa with a substitute callback function, so that the
6063 * requested block will be added to the cache.
6065 * If a read request arrives for a block that has a read in-progress,
6066 * either wait for the in-progress read to complete (and return the
6067 * results); or, if this is a read with a "done" func, add a record
6068 * to the read to invoke the "done" func when the read completes,
6069 * and return; or just return.
6071 * arc_read_done() will invoke all the requested "done" functions
6072 * for readers of this block.
6075 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
6076 arc_read_done_func_t *done, void *private, zio_priority_t priority,
6077 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
6079 arc_buf_hdr_t *hdr = NULL;
6080 kmutex_t *hash_lock = NULL;
6082 uint64_t guid = spa_load_guid(spa);
6083 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
6084 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
6085 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
6086 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
6087 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
6090 ASSERT(!BP_IS_EMBEDDED(bp) ||
6091 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
6094 if (!BP_IS_EMBEDDED(bp)) {
6096 * Embedded BP's have no DVA and require no I/O to "read".
6097 * Create an anonymous arc buf to back it.
6099 hdr = buf_hash_find(guid, bp, &hash_lock);
6103 * Determine if we have an L1 cache hit or a cache miss. For simplicity
6104 * we maintain encrypted data seperately from compressed / uncompressed
6105 * data. If the user is requesting raw encrypted data and we don't have
6106 * that in the header we will read from disk to guarantee that we can
6107 * get it even if the encryption keys aren't loaded.
6109 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
6110 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
6111 arc_buf_t *buf = NULL;
6112 *arc_flags |= ARC_FLAG_CACHED;
6114 if (HDR_IO_IN_PROGRESS(hdr)) {
6115 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
6117 ASSERT3P(head_zio, !=, NULL);
6118 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
6119 priority == ZIO_PRIORITY_SYNC_READ) {
6121 * This is a sync read that needs to wait for
6122 * an in-flight async read. Request that the
6123 * zio have its priority upgraded.
6125 zio_change_priority(head_zio, priority);
6126 DTRACE_PROBE1(arc__async__upgrade__sync,
6127 arc_buf_hdr_t *, hdr);
6128 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
6130 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6131 arc_hdr_clear_flags(hdr,
6132 ARC_FLAG_PREDICTIVE_PREFETCH);
6135 if (*arc_flags & ARC_FLAG_WAIT) {
6136 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6137 mutex_exit(hash_lock);
6140 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6143 arc_callback_t *acb = NULL;
6145 acb = kmem_zalloc(sizeof (arc_callback_t),
6147 acb->acb_done = done;
6148 acb->acb_private = private;
6149 acb->acb_compressed = compressed_read;
6150 acb->acb_encrypted = encrypted_read;
6151 acb->acb_noauth = noauth_read;
6154 acb->acb_zio_dummy = zio_null(pio,
6155 spa, NULL, NULL, NULL, zio_flags);
6157 ASSERT3P(acb->acb_done, !=, NULL);
6158 acb->acb_zio_head = head_zio;
6159 acb->acb_next = hdr->b_l1hdr.b_acb;
6160 hdr->b_l1hdr.b_acb = acb;
6161 mutex_exit(hash_lock);
6164 mutex_exit(hash_lock);
6168 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6169 hdr->b_l1hdr.b_state == arc_mfu);
6172 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6174 * This is a demand read which does not have to
6175 * wait for i/o because we did a predictive
6176 * prefetch i/o for it, which has completed.
6179 arc__demand__hit__predictive__prefetch,
6180 arc_buf_hdr_t *, hdr);
6182 arcstat_demand_hit_predictive_prefetch);
6183 arc_hdr_clear_flags(hdr,
6184 ARC_FLAG_PREDICTIVE_PREFETCH);
6187 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6189 arcstat_demand_hit_prescient_prefetch);
6190 arc_hdr_clear_flags(hdr,
6191 ARC_FLAG_PRESCIENT_PREFETCH);
6194 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
6196 /* Get a buf with the desired data in it. */
6197 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
6198 encrypted_read, compressed_read, noauth_read,
6202 * Convert authentication and decryption errors
6203 * to EIO (and generate an ereport if needed)
6204 * before leaving the ARC.
6206 rc = SET_ERROR(EIO);
6207 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6208 spa_log_error(spa, zb);
6210 FM_EREPORT_ZFS_AUTHENTICATION,
6211 spa, NULL, zb, NULL, 0, 0);
6215 (void) remove_reference(hdr, hash_lock,
6217 arc_buf_destroy_impl(buf);
6221 /* assert any errors weren't due to unloaded keys */
6222 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
6224 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6225 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
6226 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6228 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6229 arc_access(hdr, hash_lock);
6230 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6231 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6232 if (*arc_flags & ARC_FLAG_L2CACHE)
6233 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6234 mutex_exit(hash_lock);
6235 ARCSTAT_BUMP(arcstat_hits);
6236 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6237 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6238 data, metadata, hits);
6241 done(NULL, zb, bp, buf, private);
6243 uint64_t lsize = BP_GET_LSIZE(bp);
6244 uint64_t psize = BP_GET_PSIZE(bp);
6245 arc_callback_t *acb;
6248 boolean_t devw = B_FALSE;
6253 * Gracefully handle a damaged logical block size as a
6256 if (lsize > spa_maxblocksize(spa)) {
6257 rc = SET_ERROR(ECKSUM);
6262 /* this block is not in the cache */
6263 arc_buf_hdr_t *exists = NULL;
6264 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6265 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6266 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
6269 if (!BP_IS_EMBEDDED(bp)) {
6270 hdr->b_dva = *BP_IDENTITY(bp);
6271 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6272 exists = buf_hash_insert(hdr, &hash_lock);
6274 if (exists != NULL) {
6275 /* somebody beat us to the hash insert */
6276 mutex_exit(hash_lock);
6277 buf_discard_identity(hdr);
6278 arc_hdr_destroy(hdr);
6279 goto top; /* restart the IO request */
6283 * This block is in the ghost cache or encrypted data
6284 * was requested and we didn't have it. If it was
6285 * L2-only (and thus didn't have an L1 hdr),
6286 * we realloc the header to add an L1 hdr.
6288 if (!HDR_HAS_L1HDR(hdr)) {
6289 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6293 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6294 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6295 ASSERT(!HDR_HAS_RABD(hdr));
6296 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6297 ASSERT0(zfs_refcount_count(
6298 &hdr->b_l1hdr.b_refcnt));
6299 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6300 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6301 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6303 * If this header already had an IO in progress
6304 * and we are performing another IO to fetch
6305 * encrypted data we must wait until the first
6306 * IO completes so as not to confuse
6307 * arc_read_done(). This should be very rare
6308 * and so the performance impact shouldn't
6311 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6312 mutex_exit(hash_lock);
6317 * This is a delicate dance that we play here.
6318 * This hdr might be in the ghost list so we access
6319 * it to move it out of the ghost list before we
6320 * initiate the read. If it's a prefetch then
6321 * it won't have a callback so we'll remove the
6322 * reference that arc_buf_alloc_impl() created. We
6323 * do this after we've called arc_access() to
6324 * avoid hitting an assert in remove_reference().
6326 arc_access(hdr, hash_lock);
6327 arc_hdr_alloc_abd(hdr, encrypted_read);
6330 if (encrypted_read) {
6331 ASSERT(HDR_HAS_RABD(hdr));
6332 size = HDR_GET_PSIZE(hdr);
6333 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6334 zio_flags |= ZIO_FLAG_RAW;
6336 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6337 size = arc_hdr_size(hdr);
6338 hdr_abd = hdr->b_l1hdr.b_pabd;
6340 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6341 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6345 * For authenticated bp's, we do not ask the ZIO layer
6346 * to authenticate them since this will cause the entire
6347 * IO to fail if the key isn't loaded. Instead, we
6348 * defer authentication until arc_buf_fill(), which will
6349 * verify the data when the key is available.
6351 if (BP_IS_AUTHENTICATED(bp))
6352 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6355 if (*arc_flags & ARC_FLAG_PREFETCH &&
6356 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
6357 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6358 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6359 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6360 if (*arc_flags & ARC_FLAG_L2CACHE)
6361 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6362 if (BP_IS_AUTHENTICATED(bp))
6363 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6364 if (BP_GET_LEVEL(bp) > 0)
6365 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6366 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6367 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6368 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6370 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6371 acb->acb_done = done;
6372 acb->acb_private = private;
6373 acb->acb_compressed = compressed_read;
6374 acb->acb_encrypted = encrypted_read;
6375 acb->acb_noauth = noauth_read;
6378 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6379 hdr->b_l1hdr.b_acb = acb;
6380 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6382 if (HDR_HAS_L2HDR(hdr) &&
6383 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6384 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6385 addr = hdr->b_l2hdr.b_daddr;
6387 * Lock out L2ARC device removal.
6389 if (vdev_is_dead(vd) ||
6390 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6395 * We count both async reads and scrub IOs as asynchronous so
6396 * that both can be upgraded in the event of a cache hit while
6397 * the read IO is still in-flight.
6399 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6400 priority == ZIO_PRIORITY_SCRUB)
6401 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6403 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6406 * At this point, we have a level 1 cache miss. Try again in
6407 * L2ARC if possible.
6409 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6411 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
6412 uint64_t, lsize, zbookmark_phys_t *, zb);
6413 ARCSTAT_BUMP(arcstat_misses);
6414 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6415 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6416 data, metadata, misses);
6418 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6420 * Read from the L2ARC if the following are true:
6421 * 1. The L2ARC vdev was previously cached.
6422 * 2. This buffer still has L2ARC metadata.
6423 * 3. This buffer isn't currently writing to the L2ARC.
6424 * 4. The L2ARC entry wasn't evicted, which may
6425 * also have invalidated the vdev.
6426 * 5. This isn't prefetch and l2arc_noprefetch is set.
6428 if (HDR_HAS_L2HDR(hdr) &&
6429 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6430 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6431 l2arc_read_callback_t *cb;
6435 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6436 ARCSTAT_BUMP(arcstat_l2_hits);
6437 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6439 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6441 cb->l2rcb_hdr = hdr;
6444 cb->l2rcb_flags = zio_flags;
6446 asize = vdev_psize_to_asize(vd, size);
6447 if (asize != size) {
6448 abd = abd_alloc_for_io(asize,
6449 HDR_ISTYPE_METADATA(hdr));
6450 cb->l2rcb_abd = abd;
6455 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6456 addr + asize <= vd->vdev_psize -
6457 VDEV_LABEL_END_SIZE);
6460 * l2arc read. The SCL_L2ARC lock will be
6461 * released by l2arc_read_done().
6462 * Issue a null zio if the underlying buffer
6463 * was squashed to zero size by compression.
6465 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6466 ZIO_COMPRESS_EMPTY);
6467 rzio = zio_read_phys(pio, vd, addr,
6470 l2arc_read_done, cb, priority,
6471 zio_flags | ZIO_FLAG_DONT_CACHE |
6473 ZIO_FLAG_DONT_PROPAGATE |
6474 ZIO_FLAG_DONT_RETRY, B_FALSE);
6475 acb->acb_zio_head = rzio;
6477 if (hash_lock != NULL)
6478 mutex_exit(hash_lock);
6480 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6482 ARCSTAT_INCR(arcstat_l2_read_bytes,
6483 HDR_GET_PSIZE(hdr));
6485 if (*arc_flags & ARC_FLAG_NOWAIT) {
6490 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6491 if (zio_wait(rzio) == 0)
6494 /* l2arc read error; goto zio_read() */
6495 if (hash_lock != NULL)
6496 mutex_enter(hash_lock);
6498 DTRACE_PROBE1(l2arc__miss,
6499 arc_buf_hdr_t *, hdr);
6500 ARCSTAT_BUMP(arcstat_l2_misses);
6501 if (HDR_L2_WRITING(hdr))
6502 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6503 spa_config_exit(spa, SCL_L2ARC, vd);
6507 spa_config_exit(spa, SCL_L2ARC, vd);
6508 if (l2arc_ndev != 0) {
6509 DTRACE_PROBE1(l2arc__miss,
6510 arc_buf_hdr_t *, hdr);
6511 ARCSTAT_BUMP(arcstat_l2_misses);
6515 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6516 arc_read_done, hdr, priority, zio_flags, zb);
6517 acb->acb_zio_head = rzio;
6519 if (hash_lock != NULL)
6520 mutex_exit(hash_lock);
6522 if (*arc_flags & ARC_FLAG_WAIT) {
6523 rc = zio_wait(rzio);
6527 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6532 /* embedded bps don't actually go to disk */
6533 if (!BP_IS_EMBEDDED(bp))
6534 spa_read_history_add(spa, zb, *arc_flags);
6539 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6543 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6545 p->p_private = private;
6546 list_link_init(&p->p_node);
6547 zfs_refcount_create(&p->p_refcnt);
6549 mutex_enter(&arc_prune_mtx);
6550 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6551 list_insert_head(&arc_prune_list, p);
6552 mutex_exit(&arc_prune_mtx);
6558 arc_remove_prune_callback(arc_prune_t *p)
6560 boolean_t wait = B_FALSE;
6561 mutex_enter(&arc_prune_mtx);
6562 list_remove(&arc_prune_list, p);
6563 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6565 mutex_exit(&arc_prune_mtx);
6567 /* wait for arc_prune_task to finish */
6569 taskq_wait_outstanding(arc_prune_taskq, 0);
6570 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6571 zfs_refcount_destroy(&p->p_refcnt);
6572 kmem_free(p, sizeof (*p));
6576 * Notify the arc that a block was freed, and thus will never be used again.
6579 arc_freed(spa_t *spa, const blkptr_t *bp)
6582 kmutex_t *hash_lock;
6583 uint64_t guid = spa_load_guid(spa);
6585 ASSERT(!BP_IS_EMBEDDED(bp));
6587 hdr = buf_hash_find(guid, bp, &hash_lock);
6592 * We might be trying to free a block that is still doing I/O
6593 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6594 * dmu_sync-ed block). If this block is being prefetched, then it
6595 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6596 * until the I/O completes. A block may also have a reference if it is
6597 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6598 * have written the new block to its final resting place on disk but
6599 * without the dedup flag set. This would have left the hdr in the MRU
6600 * state and discoverable. When the txg finally syncs it detects that
6601 * the block was overridden in open context and issues an override I/O.
6602 * Since this is a dedup block, the override I/O will determine if the
6603 * block is already in the DDT. If so, then it will replace the io_bp
6604 * with the bp from the DDT and allow the I/O to finish. When the I/O
6605 * reaches the done callback, dbuf_write_override_done, it will
6606 * check to see if the io_bp and io_bp_override are identical.
6607 * If they are not, then it indicates that the bp was replaced with
6608 * the bp in the DDT and the override bp is freed. This allows
6609 * us to arrive here with a reference on a block that is being
6610 * freed. So if we have an I/O in progress, or a reference to
6611 * this hdr, then we don't destroy the hdr.
6613 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6614 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6615 arc_change_state(arc_anon, hdr, hash_lock);
6616 arc_hdr_destroy(hdr);
6617 mutex_exit(hash_lock);
6619 mutex_exit(hash_lock);
6625 * Release this buffer from the cache, making it an anonymous buffer. This
6626 * must be done after a read and prior to modifying the buffer contents.
6627 * If the buffer has more than one reference, we must make
6628 * a new hdr for the buffer.
6631 arc_release(arc_buf_t *buf, void *tag)
6633 arc_buf_hdr_t *hdr = buf->b_hdr;
6636 * It would be nice to assert that if its DMU metadata (level >
6637 * 0 || it's the dnode file), then it must be syncing context.
6638 * But we don't know that information at this level.
6641 mutex_enter(&buf->b_evict_lock);
6643 ASSERT(HDR_HAS_L1HDR(hdr));
6646 * We don't grab the hash lock prior to this check, because if
6647 * the buffer's header is in the arc_anon state, it won't be
6648 * linked into the hash table.
6650 if (hdr->b_l1hdr.b_state == arc_anon) {
6651 mutex_exit(&buf->b_evict_lock);
6652 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6653 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6654 ASSERT(!HDR_HAS_L2HDR(hdr));
6655 ASSERT(HDR_EMPTY(hdr));
6657 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6658 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6659 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6661 hdr->b_l1hdr.b_arc_access = 0;
6664 * If the buf is being overridden then it may already
6665 * have a hdr that is not empty.
6667 buf_discard_identity(hdr);
6673 kmutex_t *hash_lock = HDR_LOCK(hdr);
6674 mutex_enter(hash_lock);
6677 * This assignment is only valid as long as the hash_lock is
6678 * held, we must be careful not to reference state or the
6679 * b_state field after dropping the lock.
6681 arc_state_t *state = hdr->b_l1hdr.b_state;
6682 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6683 ASSERT3P(state, !=, arc_anon);
6685 /* this buffer is not on any list */
6686 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6688 if (HDR_HAS_L2HDR(hdr)) {
6689 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6692 * We have to recheck this conditional again now that
6693 * we're holding the l2ad_mtx to prevent a race with
6694 * another thread which might be concurrently calling
6695 * l2arc_evict(). In that case, l2arc_evict() might have
6696 * destroyed the header's L2 portion as we were waiting
6697 * to acquire the l2ad_mtx.
6699 if (HDR_HAS_L2HDR(hdr))
6700 arc_hdr_l2hdr_destroy(hdr);
6702 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6706 * Do we have more than one buf?
6708 if (hdr->b_l1hdr.b_bufcnt > 1) {
6709 arc_buf_hdr_t *nhdr;
6710 uint64_t spa = hdr->b_spa;
6711 uint64_t psize = HDR_GET_PSIZE(hdr);
6712 uint64_t lsize = HDR_GET_LSIZE(hdr);
6713 boolean_t protected = HDR_PROTECTED(hdr);
6714 enum zio_compress compress = arc_hdr_get_compress(hdr);
6715 arc_buf_contents_t type = arc_buf_type(hdr);
6716 VERIFY3U(hdr->b_type, ==, type);
6718 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6719 (void) remove_reference(hdr, hash_lock, tag);
6721 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6722 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6723 ASSERT(ARC_BUF_LAST(buf));
6727 * Pull the data off of this hdr and attach it to
6728 * a new anonymous hdr. Also find the last buffer
6729 * in the hdr's buffer list.
6731 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6732 ASSERT3P(lastbuf, !=, NULL);
6735 * If the current arc_buf_t and the hdr are sharing their data
6736 * buffer, then we must stop sharing that block.
6738 if (arc_buf_is_shared(buf)) {
6739 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6740 VERIFY(!arc_buf_is_shared(lastbuf));
6743 * First, sever the block sharing relationship between
6744 * buf and the arc_buf_hdr_t.
6746 arc_unshare_buf(hdr, buf);
6749 * Now we need to recreate the hdr's b_pabd. Since we
6750 * have lastbuf handy, we try to share with it, but if
6751 * we can't then we allocate a new b_pabd and copy the
6752 * data from buf into it.
6754 if (arc_can_share(hdr, lastbuf)) {
6755 arc_share_buf(hdr, lastbuf);
6757 arc_hdr_alloc_abd(hdr, B_FALSE);
6758 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6759 buf->b_data, psize);
6761 VERIFY3P(lastbuf->b_data, !=, NULL);
6762 } else if (HDR_SHARED_DATA(hdr)) {
6764 * Uncompressed shared buffers are always at the end
6765 * of the list. Compressed buffers don't have the
6766 * same requirements. This makes it hard to
6767 * simply assert that the lastbuf is shared so
6768 * we rely on the hdr's compression flags to determine
6769 * if we have a compressed, shared buffer.
6771 ASSERT(arc_buf_is_shared(lastbuf) ||
6772 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6773 ASSERT(!ARC_BUF_SHARED(buf));
6776 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6777 ASSERT3P(state, !=, arc_l2c_only);
6779 (void) zfs_refcount_remove_many(&state->arcs_size,
6780 arc_buf_size(buf), buf);
6782 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6783 ASSERT3P(state, !=, arc_l2c_only);
6784 (void) zfs_refcount_remove_many(
6785 &state->arcs_esize[type],
6786 arc_buf_size(buf), buf);
6789 hdr->b_l1hdr.b_bufcnt -= 1;
6790 if (ARC_BUF_ENCRYPTED(buf))
6791 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6793 arc_cksum_verify(buf);
6794 arc_buf_unwatch(buf);
6796 /* if this is the last uncompressed buf free the checksum */
6797 if (!arc_hdr_has_uncompressed_buf(hdr))
6798 arc_cksum_free(hdr);
6800 mutex_exit(hash_lock);
6803 * Allocate a new hdr. The new hdr will contain a b_pabd
6804 * buffer which will be freed in arc_write().
6806 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6807 compress, type, HDR_HAS_RABD(hdr));
6808 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6809 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6810 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6811 VERIFY3U(nhdr->b_type, ==, type);
6812 ASSERT(!HDR_SHARED_DATA(nhdr));
6814 nhdr->b_l1hdr.b_buf = buf;
6815 nhdr->b_l1hdr.b_bufcnt = 1;
6816 if (ARC_BUF_ENCRYPTED(buf))
6817 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6818 nhdr->b_l1hdr.b_mru_hits = 0;
6819 nhdr->b_l1hdr.b_mru_ghost_hits = 0;
6820 nhdr->b_l1hdr.b_mfu_hits = 0;
6821 nhdr->b_l1hdr.b_mfu_ghost_hits = 0;
6822 nhdr->b_l1hdr.b_l2_hits = 0;
6823 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6826 mutex_exit(&buf->b_evict_lock);
6827 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6828 HDR_GET_LSIZE(nhdr), buf);
6830 mutex_exit(&buf->b_evict_lock);
6831 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6832 /* protected by hash lock, or hdr is on arc_anon */
6833 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6834 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6835 hdr->b_l1hdr.b_mru_hits = 0;
6836 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6837 hdr->b_l1hdr.b_mfu_hits = 0;
6838 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6839 hdr->b_l1hdr.b_l2_hits = 0;
6840 arc_change_state(arc_anon, hdr, hash_lock);
6841 hdr->b_l1hdr.b_arc_access = 0;
6843 mutex_exit(hash_lock);
6844 buf_discard_identity(hdr);
6850 arc_released(arc_buf_t *buf)
6854 mutex_enter(&buf->b_evict_lock);
6855 released = (buf->b_data != NULL &&
6856 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6857 mutex_exit(&buf->b_evict_lock);
6863 arc_referenced(arc_buf_t *buf)
6867 mutex_enter(&buf->b_evict_lock);
6868 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6869 mutex_exit(&buf->b_evict_lock);
6870 return (referenced);
6875 arc_write_ready(zio_t *zio)
6877 arc_write_callback_t *callback = zio->io_private;
6878 arc_buf_t *buf = callback->awcb_buf;
6879 arc_buf_hdr_t *hdr = buf->b_hdr;
6880 blkptr_t *bp = zio->io_bp;
6881 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6882 fstrans_cookie_t cookie = spl_fstrans_mark();
6884 ASSERT(HDR_HAS_L1HDR(hdr));
6885 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6886 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6889 * If we're reexecuting this zio because the pool suspended, then
6890 * cleanup any state that was previously set the first time the
6891 * callback was invoked.
6893 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6894 arc_cksum_free(hdr);
6895 arc_buf_unwatch(buf);
6896 if (hdr->b_l1hdr.b_pabd != NULL) {
6897 if (arc_buf_is_shared(buf)) {
6898 arc_unshare_buf(hdr, buf);
6900 arc_hdr_free_abd(hdr, B_FALSE);
6904 if (HDR_HAS_RABD(hdr))
6905 arc_hdr_free_abd(hdr, B_TRUE);
6907 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6908 ASSERT(!HDR_HAS_RABD(hdr));
6909 ASSERT(!HDR_SHARED_DATA(hdr));
6910 ASSERT(!arc_buf_is_shared(buf));
6912 callback->awcb_ready(zio, buf, callback->awcb_private);
6914 if (HDR_IO_IN_PROGRESS(hdr))
6915 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6917 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6919 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6920 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6922 if (BP_IS_PROTECTED(bp)) {
6923 /* ZIL blocks are written through zio_rewrite */
6924 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6925 ASSERT(HDR_PROTECTED(hdr));
6927 if (BP_SHOULD_BYTESWAP(bp)) {
6928 if (BP_GET_LEVEL(bp) > 0) {
6929 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6931 hdr->b_l1hdr.b_byteswap =
6932 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6935 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6938 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6939 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6940 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6941 hdr->b_crypt_hdr.b_iv);
6942 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6946 * If this block was written for raw encryption but the zio layer
6947 * ended up only authenticating it, adjust the buffer flags now.
6949 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6950 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6951 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6952 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6953 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6954 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6955 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6956 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6959 /* this must be done after the buffer flags are adjusted */
6960 arc_cksum_compute(buf);
6962 enum zio_compress compress;
6963 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6964 compress = ZIO_COMPRESS_OFF;
6966 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6967 compress = BP_GET_COMPRESS(bp);
6969 HDR_SET_PSIZE(hdr, psize);
6970 arc_hdr_set_compress(hdr, compress);
6972 if (zio->io_error != 0 || psize == 0)
6976 * Fill the hdr with data. If the buffer is encrypted we have no choice
6977 * but to copy the data into b_radb. If the hdr is compressed, the data
6978 * we want is available from the zio, otherwise we can take it from
6981 * We might be able to share the buf's data with the hdr here. However,
6982 * doing so would cause the ARC to be full of linear ABDs if we write a
6983 * lot of shareable data. As a compromise, we check whether scattered
6984 * ABDs are allowed, and assume that if they are then the user wants
6985 * the ARC to be primarily filled with them regardless of the data being
6986 * written. Therefore, if they're allowed then we allocate one and copy
6987 * the data into it; otherwise, we share the data directly if we can.
6989 if (ARC_BUF_ENCRYPTED(buf)) {
6990 ASSERT3U(psize, >, 0);
6991 ASSERT(ARC_BUF_COMPRESSED(buf));
6992 arc_hdr_alloc_abd(hdr, B_TRUE);
6993 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6994 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6996 * Ideally, we would always copy the io_abd into b_pabd, but the
6997 * user may have disabled compressed ARC, thus we must check the
6998 * hdr's compression setting rather than the io_bp's.
7000 if (BP_IS_ENCRYPTED(bp)) {
7001 ASSERT3U(psize, >, 0);
7002 arc_hdr_alloc_abd(hdr, B_TRUE);
7003 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
7004 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
7005 !ARC_BUF_COMPRESSED(buf)) {
7006 ASSERT3U(psize, >, 0);
7007 arc_hdr_alloc_abd(hdr, B_FALSE);
7008 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
7010 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
7011 arc_hdr_alloc_abd(hdr, B_FALSE);
7012 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
7016 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
7017 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
7018 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
7020 arc_share_buf(hdr, buf);
7024 arc_hdr_verify(hdr, bp);
7025 spl_fstrans_unmark(cookie);
7029 arc_write_children_ready(zio_t *zio)
7031 arc_write_callback_t *callback = zio->io_private;
7032 arc_buf_t *buf = callback->awcb_buf;
7034 callback->awcb_children_ready(zio, buf, callback->awcb_private);
7038 * The SPA calls this callback for each physical write that happens on behalf
7039 * of a logical write. See the comment in dbuf_write_physdone() for details.
7042 arc_write_physdone(zio_t *zio)
7044 arc_write_callback_t *cb = zio->io_private;
7045 if (cb->awcb_physdone != NULL)
7046 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
7050 arc_write_done(zio_t *zio)
7052 arc_write_callback_t *callback = zio->io_private;
7053 arc_buf_t *buf = callback->awcb_buf;
7054 arc_buf_hdr_t *hdr = buf->b_hdr;
7056 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7058 if (zio->io_error == 0) {
7059 arc_hdr_verify(hdr, zio->io_bp);
7061 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
7062 buf_discard_identity(hdr);
7064 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
7065 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
7068 ASSERT(HDR_EMPTY(hdr));
7072 * If the block to be written was all-zero or compressed enough to be
7073 * embedded in the BP, no write was performed so there will be no
7074 * dva/birth/checksum. The buffer must therefore remain anonymous
7077 if (!HDR_EMPTY(hdr)) {
7078 arc_buf_hdr_t *exists;
7079 kmutex_t *hash_lock;
7081 ASSERT3U(zio->io_error, ==, 0);
7083 arc_cksum_verify(buf);
7085 exists = buf_hash_insert(hdr, &hash_lock);
7086 if (exists != NULL) {
7088 * This can only happen if we overwrite for
7089 * sync-to-convergence, because we remove
7090 * buffers from the hash table when we arc_free().
7092 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
7093 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7094 panic("bad overwrite, hdr=%p exists=%p",
7095 (void *)hdr, (void *)exists);
7096 ASSERT(zfs_refcount_is_zero(
7097 &exists->b_l1hdr.b_refcnt));
7098 arc_change_state(arc_anon, exists, hash_lock);
7099 mutex_exit(hash_lock);
7100 arc_hdr_destroy(exists);
7101 exists = buf_hash_insert(hdr, &hash_lock);
7102 ASSERT3P(exists, ==, NULL);
7103 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
7105 ASSERT(zio->io_prop.zp_nopwrite);
7106 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7107 panic("bad nopwrite, hdr=%p exists=%p",
7108 (void *)hdr, (void *)exists);
7111 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
7112 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
7113 ASSERT(BP_GET_DEDUP(zio->io_bp));
7114 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
7117 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7118 /* if it's not anon, we are doing a scrub */
7119 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
7120 arc_access(hdr, hash_lock);
7121 mutex_exit(hash_lock);
7123 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7126 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
7127 callback->awcb_done(zio, buf, callback->awcb_private);
7129 abd_put(zio->io_abd);
7130 kmem_free(callback, sizeof (arc_write_callback_t));
7134 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
7135 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
7136 const zio_prop_t *zp, arc_write_done_func_t *ready,
7137 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
7138 arc_write_done_func_t *done, void *private, zio_priority_t priority,
7139 int zio_flags, const zbookmark_phys_t *zb)
7141 arc_buf_hdr_t *hdr = buf->b_hdr;
7142 arc_write_callback_t *callback;
7144 zio_prop_t localprop = *zp;
7146 ASSERT3P(ready, !=, NULL);
7147 ASSERT3P(done, !=, NULL);
7148 ASSERT(!HDR_IO_ERROR(hdr));
7149 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
7150 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7151 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
7153 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
7155 if (ARC_BUF_ENCRYPTED(buf)) {
7156 ASSERT(ARC_BUF_COMPRESSED(buf));
7157 localprop.zp_encrypt = B_TRUE;
7158 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7159 localprop.zp_byteorder =
7160 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
7161 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
7162 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
7164 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
7166 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
7168 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
7169 localprop.zp_nopwrite = B_FALSE;
7170 localprop.zp_copies =
7171 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
7173 zio_flags |= ZIO_FLAG_RAW;
7174 } else if (ARC_BUF_COMPRESSED(buf)) {
7175 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7176 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7177 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7179 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7180 callback->awcb_ready = ready;
7181 callback->awcb_children_ready = children_ready;
7182 callback->awcb_physdone = physdone;
7183 callback->awcb_done = done;
7184 callback->awcb_private = private;
7185 callback->awcb_buf = buf;
7188 * The hdr's b_pabd is now stale, free it now. A new data block
7189 * will be allocated when the zio pipeline calls arc_write_ready().
7191 if (hdr->b_l1hdr.b_pabd != NULL) {
7193 * If the buf is currently sharing the data block with
7194 * the hdr then we need to break that relationship here.
7195 * The hdr will remain with a NULL data pointer and the
7196 * buf will take sole ownership of the block.
7198 if (arc_buf_is_shared(buf)) {
7199 arc_unshare_buf(hdr, buf);
7201 arc_hdr_free_abd(hdr, B_FALSE);
7203 VERIFY3P(buf->b_data, !=, NULL);
7206 if (HDR_HAS_RABD(hdr))
7207 arc_hdr_free_abd(hdr, B_TRUE);
7209 if (!(zio_flags & ZIO_FLAG_RAW))
7210 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7212 ASSERT(!arc_buf_is_shared(buf));
7213 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7215 zio = zio_write(pio, spa, txg, bp,
7216 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7217 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7218 (children_ready != NULL) ? arc_write_children_ready : NULL,
7219 arc_write_physdone, arc_write_done, callback,
7220 priority, zio_flags, zb);
7226 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
7229 uint64_t available_memory = arc_free_memory();
7233 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
7236 if (available_memory > arc_all_memory() * arc_lotsfree_percent / 100)
7239 if (txg > spa->spa_lowmem_last_txg) {
7240 spa->spa_lowmem_last_txg = txg;
7241 spa->spa_lowmem_page_load = 0;
7244 * If we are in pageout, we know that memory is already tight,
7245 * the arc is already going to be evicting, so we just want to
7246 * continue to let page writes occur as quickly as possible.
7248 if (current_is_kswapd()) {
7249 if (spa->spa_lowmem_page_load >
7250 MAX(arc_sys_free / 4, available_memory) / 4) {
7251 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7252 return (SET_ERROR(ERESTART));
7254 /* Note: reserve is inflated, so we deflate */
7255 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
7257 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
7258 /* memory is low, delay before restarting */
7259 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
7260 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7261 return (SET_ERROR(EAGAIN));
7263 spa->spa_lowmem_page_load = 0;
7264 #endif /* _KERNEL */
7269 arc_tempreserve_clear(uint64_t reserve)
7271 atomic_add_64(&arc_tempreserve, -reserve);
7272 ASSERT((int64_t)arc_tempreserve >= 0);
7276 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7282 reserve > arc_c/4 &&
7283 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7284 arc_c = MIN(arc_c_max, reserve * 4);
7287 * Throttle when the calculated memory footprint for the TXG
7288 * exceeds the target ARC size.
7290 if (reserve > arc_c) {
7291 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7292 return (SET_ERROR(ERESTART));
7296 * Don't count loaned bufs as in flight dirty data to prevent long
7297 * network delays from blocking transactions that are ready to be
7298 * assigned to a txg.
7301 /* assert that it has not wrapped around */
7302 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7304 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7305 arc_loaned_bytes), 0);
7308 * Writes will, almost always, require additional memory allocations
7309 * in order to compress/encrypt/etc the data. We therefore need to
7310 * make sure that there is sufficient available memory for this.
7312 error = arc_memory_throttle(spa, reserve, txg);
7317 * Throttle writes when the amount of dirty data in the cache
7318 * gets too large. We try to keep the cache less than half full
7319 * of dirty blocks so that our sync times don't grow too large.
7321 * In the case of one pool being built on another pool, we want
7322 * to make sure we don't end up throttling the lower (backing)
7323 * pool when the upper pool is the majority contributor to dirty
7324 * data. To insure we make forward progress during throttling, we
7325 * also check the current pool's net dirty data and only throttle
7326 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7327 * data in the cache.
7329 * Note: if two requests come in concurrently, we might let them
7330 * both succeed, when one of them should fail. Not a huge deal.
7332 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7333 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7335 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
7336 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
7337 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7339 uint64_t meta_esize = zfs_refcount_count(
7340 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7341 uint64_t data_esize =
7342 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7343 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7344 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7345 arc_tempreserve >> 10, meta_esize >> 10,
7346 data_esize >> 10, reserve >> 10, arc_c >> 10);
7348 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7349 return (SET_ERROR(ERESTART));
7351 atomic_add_64(&arc_tempreserve, reserve);
7356 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7357 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7359 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7360 evict_data->value.ui64 =
7361 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7362 evict_metadata->value.ui64 =
7363 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7367 arc_kstat_update(kstat_t *ksp, int rw)
7369 arc_stats_t *as = ksp->ks_data;
7371 if (rw == KSTAT_WRITE) {
7372 return (SET_ERROR(EACCES));
7374 arc_kstat_update_state(arc_anon,
7375 &as->arcstat_anon_size,
7376 &as->arcstat_anon_evictable_data,
7377 &as->arcstat_anon_evictable_metadata);
7378 arc_kstat_update_state(arc_mru,
7379 &as->arcstat_mru_size,
7380 &as->arcstat_mru_evictable_data,
7381 &as->arcstat_mru_evictable_metadata);
7382 arc_kstat_update_state(arc_mru_ghost,
7383 &as->arcstat_mru_ghost_size,
7384 &as->arcstat_mru_ghost_evictable_data,
7385 &as->arcstat_mru_ghost_evictable_metadata);
7386 arc_kstat_update_state(arc_mfu,
7387 &as->arcstat_mfu_size,
7388 &as->arcstat_mfu_evictable_data,
7389 &as->arcstat_mfu_evictable_metadata);
7390 arc_kstat_update_state(arc_mfu_ghost,
7391 &as->arcstat_mfu_ghost_size,
7392 &as->arcstat_mfu_ghost_evictable_data,
7393 &as->arcstat_mfu_ghost_evictable_metadata);
7395 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
7396 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
7397 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
7398 ARCSTAT(arcstat_metadata_size) =
7399 aggsum_value(&astat_metadata_size);
7400 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
7401 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
7402 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
7403 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
7404 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
7406 as->arcstat_memory_all_bytes.value.ui64 =
7408 as->arcstat_memory_free_bytes.value.ui64 =
7410 as->arcstat_memory_available_bytes.value.i64 =
7411 arc_available_memory();
7418 * This function *must* return indices evenly distributed between all
7419 * sublists of the multilist. This is needed due to how the ARC eviction
7420 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7421 * distributed between all sublists and uses this assumption when
7422 * deciding which sublist to evict from and how much to evict from it.
7425 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7427 arc_buf_hdr_t *hdr = obj;
7430 * We rely on b_dva to generate evenly distributed index
7431 * numbers using buf_hash below. So, as an added precaution,
7432 * let's make sure we never add empty buffers to the arc lists.
7434 ASSERT(!HDR_EMPTY(hdr));
7437 * The assumption here, is the hash value for a given
7438 * arc_buf_hdr_t will remain constant throughout its lifetime
7439 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7440 * Thus, we don't need to store the header's sublist index
7441 * on insertion, as this index can be recalculated on removal.
7443 * Also, the low order bits of the hash value are thought to be
7444 * distributed evenly. Otherwise, in the case that the multilist
7445 * has a power of two number of sublists, each sublists' usage
7446 * would not be evenly distributed.
7448 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7449 multilist_get_num_sublists(ml));
7453 * Called during module initialization and periodically thereafter to
7454 * apply reasonable changes to the exposed performance tunings. Non-zero
7455 * zfs_* values which differ from the currently set values will be applied.
7458 arc_tuning_update(void)
7460 uint64_t allmem = arc_all_memory();
7461 unsigned long limit;
7463 /* Valid range: 64M - <all physical memory> */
7464 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7465 (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) &&
7466 (zfs_arc_max > arc_c_min)) {
7467 arc_c_max = zfs_arc_max;
7469 arc_p = (arc_c >> 1);
7470 if (arc_meta_limit > arc_c_max)
7471 arc_meta_limit = arc_c_max;
7472 if (arc_dnode_limit > arc_meta_limit)
7473 arc_dnode_limit = arc_meta_limit;
7476 /* Valid range: 32M - <arc_c_max> */
7477 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7478 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7479 (zfs_arc_min <= arc_c_max)) {
7480 arc_c_min = zfs_arc_min;
7481 arc_c = MAX(arc_c, arc_c_min);
7484 /* Valid range: 16M - <arc_c_max> */
7485 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7486 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7487 (zfs_arc_meta_min <= arc_c_max)) {
7488 arc_meta_min = zfs_arc_meta_min;
7489 if (arc_meta_limit < arc_meta_min)
7490 arc_meta_limit = arc_meta_min;
7491 if (arc_dnode_limit < arc_meta_min)
7492 arc_dnode_limit = arc_meta_min;
7495 /* Valid range: <arc_meta_min> - <arc_c_max> */
7496 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7497 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7498 if ((limit != arc_meta_limit) &&
7499 (limit >= arc_meta_min) &&
7500 (limit <= arc_c_max))
7501 arc_meta_limit = limit;
7503 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7504 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7505 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7506 if ((limit != arc_dnode_limit) &&
7507 (limit >= arc_meta_min) &&
7508 (limit <= arc_meta_limit))
7509 arc_dnode_limit = limit;
7511 /* Valid range: 1 - N */
7512 if (zfs_arc_grow_retry)
7513 arc_grow_retry = zfs_arc_grow_retry;
7515 /* Valid range: 1 - N */
7516 if (zfs_arc_shrink_shift) {
7517 arc_shrink_shift = zfs_arc_shrink_shift;
7518 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7521 /* Valid range: 1 - N */
7522 if (zfs_arc_p_min_shift)
7523 arc_p_min_shift = zfs_arc_p_min_shift;
7525 /* Valid range: 1 - N ms */
7526 if (zfs_arc_min_prefetch_ms)
7527 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7529 /* Valid range: 1 - N ms */
7530 if (zfs_arc_min_prescient_prefetch_ms) {
7531 arc_min_prescient_prefetch_ms =
7532 zfs_arc_min_prescient_prefetch_ms;
7535 /* Valid range: 0 - 100 */
7536 if ((zfs_arc_lotsfree_percent >= 0) &&
7537 (zfs_arc_lotsfree_percent <= 100))
7538 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7540 /* Valid range: 0 - <all physical memory> */
7541 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7542 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7547 arc_state_init(void)
7549 arc_anon = &ARC_anon;
7551 arc_mru_ghost = &ARC_mru_ghost;
7553 arc_mfu_ghost = &ARC_mfu_ghost;
7554 arc_l2c_only = &ARC_l2c_only;
7556 arc_mru->arcs_list[ARC_BUFC_METADATA] =
7557 multilist_create(sizeof (arc_buf_hdr_t),
7558 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7559 arc_state_multilist_index_func);
7560 arc_mru->arcs_list[ARC_BUFC_DATA] =
7561 multilist_create(sizeof (arc_buf_hdr_t),
7562 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7563 arc_state_multilist_index_func);
7564 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
7565 multilist_create(sizeof (arc_buf_hdr_t),
7566 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7567 arc_state_multilist_index_func);
7568 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7569 multilist_create(sizeof (arc_buf_hdr_t),
7570 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7571 arc_state_multilist_index_func);
7572 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7573 multilist_create(sizeof (arc_buf_hdr_t),
7574 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7575 arc_state_multilist_index_func);
7576 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7577 multilist_create(sizeof (arc_buf_hdr_t),
7578 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7579 arc_state_multilist_index_func);
7580 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7581 multilist_create(sizeof (arc_buf_hdr_t),
7582 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7583 arc_state_multilist_index_func);
7584 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7585 multilist_create(sizeof (arc_buf_hdr_t),
7586 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7587 arc_state_multilist_index_func);
7588 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7589 multilist_create(sizeof (arc_buf_hdr_t),
7590 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7591 arc_state_multilist_index_func);
7592 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7593 multilist_create(sizeof (arc_buf_hdr_t),
7594 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7595 arc_state_multilist_index_func);
7597 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7598 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7599 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7600 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7601 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7602 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7603 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7604 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7605 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7606 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7607 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7608 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7610 zfs_refcount_create(&arc_anon->arcs_size);
7611 zfs_refcount_create(&arc_mru->arcs_size);
7612 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7613 zfs_refcount_create(&arc_mfu->arcs_size);
7614 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7615 zfs_refcount_create(&arc_l2c_only->arcs_size);
7617 aggsum_init(&arc_meta_used, 0);
7618 aggsum_init(&arc_size, 0);
7619 aggsum_init(&astat_data_size, 0);
7620 aggsum_init(&astat_metadata_size, 0);
7621 aggsum_init(&astat_hdr_size, 0);
7622 aggsum_init(&astat_l2_hdr_size, 0);
7623 aggsum_init(&astat_bonus_size, 0);
7624 aggsum_init(&astat_dnode_size, 0);
7625 aggsum_init(&astat_dbuf_size, 0);
7627 arc_anon->arcs_state = ARC_STATE_ANON;
7628 arc_mru->arcs_state = ARC_STATE_MRU;
7629 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7630 arc_mfu->arcs_state = ARC_STATE_MFU;
7631 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7632 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7636 arc_state_fini(void)
7638 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7639 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7640 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7641 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7642 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7643 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7644 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7645 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7646 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7647 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7648 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7649 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7651 zfs_refcount_destroy(&arc_anon->arcs_size);
7652 zfs_refcount_destroy(&arc_mru->arcs_size);
7653 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7654 zfs_refcount_destroy(&arc_mfu->arcs_size);
7655 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7656 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7658 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7659 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7660 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7661 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7662 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7663 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7664 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7665 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7666 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7667 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7669 aggsum_fini(&arc_meta_used);
7670 aggsum_fini(&arc_size);
7671 aggsum_fini(&astat_data_size);
7672 aggsum_fini(&astat_metadata_size);
7673 aggsum_fini(&astat_hdr_size);
7674 aggsum_fini(&astat_l2_hdr_size);
7675 aggsum_fini(&astat_bonus_size);
7676 aggsum_fini(&astat_dnode_size);
7677 aggsum_fini(&astat_dbuf_size);
7681 arc_target_bytes(void)
7689 uint64_t percent, allmem = arc_all_memory();
7691 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
7692 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
7693 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
7695 arc_min_prefetch_ms = 1000;
7696 arc_min_prescient_prefetch_ms = 6000;
7700 * Register a shrinker to support synchronous (direct) memory
7701 * reclaim from the arc. This is done to prevent kswapd from
7702 * swapping out pages when it is preferable to shrink the arc.
7704 spl_register_shrinker(&arc_shrinker);
7706 /* Set to 1/64 of all memory or a minimum of 512K */
7707 arc_sys_free = MAX(allmem / 64, (512 * 1024));
7711 /* Set max to 1/2 of all memory */
7712 arc_c_max = allmem / 2;
7715 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
7716 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7719 * In userland, there's only the memory pressure that we artificially
7720 * create (see arc_available_memory()). Don't let arc_c get too
7721 * small, because it can cause transactions to be larger than
7722 * arc_c, causing arc_tempreserve_space() to fail.
7724 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7728 arc_p = (arc_c >> 1);
7730 /* Set min to 1/2 of arc_c_min */
7731 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7732 /* Initialize maximum observed usage to zero */
7735 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7736 * arc_meta_min, and a ceiling of arc_c_max.
7738 percent = MIN(zfs_arc_meta_limit_percent, 100);
7739 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7740 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7741 arc_dnode_limit = (percent * arc_meta_limit) / 100;
7743 /* Apply user specified tunings */
7744 arc_tuning_update();
7746 /* if kmem_flags are set, lets try to use less memory */
7747 if (kmem_debugging())
7749 if (arc_c < arc_c_min)
7755 list_create(&arc_prune_list, sizeof (arc_prune_t),
7756 offsetof(arc_prune_t, p_node));
7757 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7759 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, defclsyspri,
7760 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7762 arc_reclaim_thread_exit = B_FALSE;
7764 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7765 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7767 if (arc_ksp != NULL) {
7768 arc_ksp->ks_data = &arc_stats;
7769 arc_ksp->ks_update = arc_kstat_update;
7770 kstat_install(arc_ksp);
7773 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
7774 TS_RUN, defclsyspri);
7780 * Calculate maximum amount of dirty data per pool.
7782 * If it has been set by a module parameter, take that.
7783 * Otherwise, use a percentage of physical memory defined by
7784 * zfs_dirty_data_max_percent (default 10%) with a cap at
7785 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7787 if (zfs_dirty_data_max_max == 0)
7788 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7789 allmem * zfs_dirty_data_max_max_percent / 100);
7791 if (zfs_dirty_data_max == 0) {
7792 zfs_dirty_data_max = allmem *
7793 zfs_dirty_data_max_percent / 100;
7794 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7795 zfs_dirty_data_max_max);
7805 spl_unregister_shrinker(&arc_shrinker);
7806 #endif /* _KERNEL */
7808 mutex_enter(&arc_reclaim_lock);
7809 arc_reclaim_thread_exit = B_TRUE;
7811 * The reclaim thread will set arc_reclaim_thread_exit back to
7812 * B_FALSE when it is finished exiting; we're waiting for that.
7814 while (arc_reclaim_thread_exit) {
7815 cv_signal(&arc_reclaim_thread_cv);
7816 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
7818 mutex_exit(&arc_reclaim_lock);
7820 /* Use B_TRUE to ensure *all* buffers are evicted */
7821 arc_flush(NULL, B_TRUE);
7825 if (arc_ksp != NULL) {
7826 kstat_delete(arc_ksp);
7830 taskq_wait(arc_prune_taskq);
7831 taskq_destroy(arc_prune_taskq);
7833 mutex_enter(&arc_prune_mtx);
7834 while ((p = list_head(&arc_prune_list)) != NULL) {
7835 list_remove(&arc_prune_list, p);
7836 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7837 zfs_refcount_destroy(&p->p_refcnt);
7838 kmem_free(p, sizeof (*p));
7840 mutex_exit(&arc_prune_mtx);
7842 list_destroy(&arc_prune_list);
7843 mutex_destroy(&arc_prune_mtx);
7844 mutex_destroy(&arc_reclaim_lock);
7845 cv_destroy(&arc_reclaim_thread_cv);
7846 cv_destroy(&arc_reclaim_waiters_cv);
7851 ASSERT0(arc_loaned_bytes);
7857 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7858 * It uses dedicated storage devices to hold cached data, which are populated
7859 * using large infrequent writes. The main role of this cache is to boost
7860 * the performance of random read workloads. The intended L2ARC devices
7861 * include short-stroked disks, solid state disks, and other media with
7862 * substantially faster read latency than disk.
7864 * +-----------------------+
7866 * +-----------------------+
7869 * l2arc_feed_thread() arc_read()
7873 * +---------------+ |
7875 * +---------------+ |
7880 * +-------+ +-------+
7882 * | cache | | cache |
7883 * +-------+ +-------+
7884 * +=========+ .-----.
7885 * : L2ARC : |-_____-|
7886 * : devices : | Disks |
7887 * +=========+ `-_____-'
7889 * Read requests are satisfied from the following sources, in order:
7892 * 2) vdev cache of L2ARC devices
7894 * 4) vdev cache of disks
7897 * Some L2ARC device types exhibit extremely slow write performance.
7898 * To accommodate for this there are some significant differences between
7899 * the L2ARC and traditional cache design:
7901 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7902 * the ARC behave as usual, freeing buffers and placing headers on ghost
7903 * lists. The ARC does not send buffers to the L2ARC during eviction as
7904 * this would add inflated write latencies for all ARC memory pressure.
7906 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7907 * It does this by periodically scanning buffers from the eviction-end of
7908 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7909 * not already there. It scans until a headroom of buffers is satisfied,
7910 * which itself is a buffer for ARC eviction. If a compressible buffer is
7911 * found during scanning and selected for writing to an L2ARC device, we
7912 * temporarily boost scanning headroom during the next scan cycle to make
7913 * sure we adapt to compression effects (which might significantly reduce
7914 * the data volume we write to L2ARC). The thread that does this is
7915 * l2arc_feed_thread(), illustrated below; example sizes are included to
7916 * provide a better sense of ratio than this diagram:
7919 * +---------------------+----------+
7920 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7921 * +---------------------+----------+ | o L2ARC eligible
7922 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7923 * +---------------------+----------+ |
7924 * 15.9 Gbytes ^ 32 Mbytes |
7926 * l2arc_feed_thread()
7928 * l2arc write hand <--[oooo]--'
7932 * +==============================+
7933 * L2ARC dev |####|#|###|###| |####| ... |
7934 * +==============================+
7937 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7938 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7939 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7940 * safe to say that this is an uncommon case, since buffers at the end of
7941 * the ARC lists have moved there due to inactivity.
7943 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7944 * then the L2ARC simply misses copying some buffers. This serves as a
7945 * pressure valve to prevent heavy read workloads from both stalling the ARC
7946 * with waits and clogging the L2ARC with writes. This also helps prevent
7947 * the potential for the L2ARC to churn if it attempts to cache content too
7948 * quickly, such as during backups of the entire pool.
7950 * 5. After system boot and before the ARC has filled main memory, there are
7951 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7952 * lists can remain mostly static. Instead of searching from tail of these
7953 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7954 * for eligible buffers, greatly increasing its chance of finding them.
7956 * The L2ARC device write speed is also boosted during this time so that
7957 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7958 * there are no L2ARC reads, and no fear of degrading read performance
7959 * through increased writes.
7961 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7962 * the vdev queue can aggregate them into larger and fewer writes. Each
7963 * device is written to in a rotor fashion, sweeping writes through
7964 * available space then repeating.
7966 * 7. The L2ARC does not store dirty content. It never needs to flush
7967 * write buffers back to disk based storage.
7969 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7970 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7972 * The performance of the L2ARC can be tweaked by a number of tunables, which
7973 * may be necessary for different workloads:
7975 * l2arc_write_max max write bytes per interval
7976 * l2arc_write_boost extra write bytes during device warmup
7977 * l2arc_noprefetch skip caching prefetched buffers
7978 * l2arc_headroom number of max device writes to precache
7979 * l2arc_headroom_boost when we find compressed buffers during ARC
7980 * scanning, we multiply headroom by this
7981 * percentage factor for the next scan cycle,
7982 * since more compressed buffers are likely to
7984 * l2arc_feed_secs seconds between L2ARC writing
7986 * Tunables may be removed or added as future performance improvements are
7987 * integrated, and also may become zpool properties.
7989 * There are three key functions that control how the L2ARC warms up:
7991 * l2arc_write_eligible() check if a buffer is eligible to cache
7992 * l2arc_write_size() calculate how much to write
7993 * l2arc_write_interval() calculate sleep delay between writes
7995 * These three functions determine what to write, how much, and how quickly
8000 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8003 * A buffer is *not* eligible for the L2ARC if it:
8004 * 1. belongs to a different spa.
8005 * 2. is already cached on the L2ARC.
8006 * 3. has an I/O in progress (it may be an incomplete read).
8007 * 4. is flagged not eligible (zfs property).
8009 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8010 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8017 l2arc_write_size(void)
8022 * Make sure our globals have meaningful values in case the user
8025 size = l2arc_write_max;
8027 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8028 "be greater than zero, resetting it to the default (%d)",
8030 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8033 if (arc_warm == B_FALSE)
8034 size += l2arc_write_boost;
8041 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8043 clock_t interval, next, now;
8046 * If the ARC lists are busy, increase our write rate; if the
8047 * lists are stale, idle back. This is achieved by checking
8048 * how much we previously wrote - if it was more than half of
8049 * what we wanted, schedule the next write much sooner.
8051 if (l2arc_feed_again && wrote > (wanted / 2))
8052 interval = (hz * l2arc_feed_min_ms) / 1000;
8054 interval = hz * l2arc_feed_secs;
8056 now = ddi_get_lbolt();
8057 next = MAX(now, MIN(now + interval, began + interval));
8063 * Cycle through L2ARC devices. This is how L2ARC load balances.
8064 * If a device is returned, this also returns holding the spa config lock.
8066 static l2arc_dev_t *
8067 l2arc_dev_get_next(void)
8069 l2arc_dev_t *first, *next = NULL;
8072 * Lock out the removal of spas (spa_namespace_lock), then removal
8073 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8074 * both locks will be dropped and a spa config lock held instead.
8076 mutex_enter(&spa_namespace_lock);
8077 mutex_enter(&l2arc_dev_mtx);
8079 /* if there are no vdevs, there is nothing to do */
8080 if (l2arc_ndev == 0)
8084 next = l2arc_dev_last;
8086 /* loop around the list looking for a non-faulted vdev */
8088 next = list_head(l2arc_dev_list);
8090 next = list_next(l2arc_dev_list, next);
8092 next = list_head(l2arc_dev_list);
8095 /* if we have come back to the start, bail out */
8098 else if (next == first)
8101 } while (vdev_is_dead(next->l2ad_vdev));
8103 /* if we were unable to find any usable vdevs, return NULL */
8104 if (vdev_is_dead(next->l2ad_vdev))
8107 l2arc_dev_last = next;
8110 mutex_exit(&l2arc_dev_mtx);
8113 * Grab the config lock to prevent the 'next' device from being
8114 * removed while we are writing to it.
8117 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8118 mutex_exit(&spa_namespace_lock);
8124 * Free buffers that were tagged for destruction.
8127 l2arc_do_free_on_write(void)
8130 l2arc_data_free_t *df, *df_prev;
8132 mutex_enter(&l2arc_free_on_write_mtx);
8133 buflist = l2arc_free_on_write;
8135 for (df = list_tail(buflist); df; df = df_prev) {
8136 df_prev = list_prev(buflist, df);
8137 ASSERT3P(df->l2df_abd, !=, NULL);
8138 abd_free(df->l2df_abd);
8139 list_remove(buflist, df);
8140 kmem_free(df, sizeof (l2arc_data_free_t));
8143 mutex_exit(&l2arc_free_on_write_mtx);
8147 * A write to a cache device has completed. Update all headers to allow
8148 * reads from these buffers to begin.
8151 l2arc_write_done(zio_t *zio)
8153 l2arc_write_callback_t *cb;
8156 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8157 kmutex_t *hash_lock;
8158 int64_t bytes_dropped = 0;
8160 cb = zio->io_private;
8161 ASSERT3P(cb, !=, NULL);
8162 dev = cb->l2wcb_dev;
8163 ASSERT3P(dev, !=, NULL);
8164 head = cb->l2wcb_head;
8165 ASSERT3P(head, !=, NULL);
8166 buflist = &dev->l2ad_buflist;
8167 ASSERT3P(buflist, !=, NULL);
8168 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8169 l2arc_write_callback_t *, cb);
8171 if (zio->io_error != 0)
8172 ARCSTAT_BUMP(arcstat_l2_writes_error);
8175 * All writes completed, or an error was hit.
8178 mutex_enter(&dev->l2ad_mtx);
8179 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8180 hdr_prev = list_prev(buflist, hdr);
8182 hash_lock = HDR_LOCK(hdr);
8185 * We cannot use mutex_enter or else we can deadlock
8186 * with l2arc_write_buffers (due to swapping the order
8187 * the hash lock and l2ad_mtx are taken).
8189 if (!mutex_tryenter(hash_lock)) {
8191 * Missed the hash lock. We must retry so we
8192 * don't leave the ARC_FLAG_L2_WRITING bit set.
8194 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8197 * We don't want to rescan the headers we've
8198 * already marked as having been written out, so
8199 * we reinsert the head node so we can pick up
8200 * where we left off.
8202 list_remove(buflist, head);
8203 list_insert_after(buflist, hdr, head);
8205 mutex_exit(&dev->l2ad_mtx);
8208 * We wait for the hash lock to become available
8209 * to try and prevent busy waiting, and increase
8210 * the chance we'll be able to acquire the lock
8211 * the next time around.
8213 mutex_enter(hash_lock);
8214 mutex_exit(hash_lock);
8219 * We could not have been moved into the arc_l2c_only
8220 * state while in-flight due to our ARC_FLAG_L2_WRITING
8221 * bit being set. Let's just ensure that's being enforced.
8223 ASSERT(HDR_HAS_L1HDR(hdr));
8226 * Skipped - drop L2ARC entry and mark the header as no
8227 * longer L2 eligibile.
8229 if (zio->io_error != 0) {
8231 * Error - drop L2ARC entry.
8233 list_remove(buflist, hdr);
8234 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8236 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
8237 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
8239 bytes_dropped += arc_hdr_size(hdr);
8240 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8241 arc_hdr_size(hdr), hdr);
8245 * Allow ARC to begin reads and ghost list evictions to
8248 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8250 mutex_exit(hash_lock);
8253 atomic_inc_64(&l2arc_writes_done);
8254 list_remove(buflist, head);
8255 ASSERT(!HDR_HAS_L1HDR(head));
8256 kmem_cache_free(hdr_l2only_cache, head);
8257 mutex_exit(&dev->l2ad_mtx);
8259 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8261 l2arc_do_free_on_write();
8263 kmem_free(cb, sizeof (l2arc_write_callback_t));
8267 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8270 spa_t *spa = zio->io_spa;
8271 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8272 blkptr_t *bp = zio->io_bp;
8273 uint8_t salt[ZIO_DATA_SALT_LEN];
8274 uint8_t iv[ZIO_DATA_IV_LEN];
8275 uint8_t mac[ZIO_DATA_MAC_LEN];
8276 boolean_t no_crypt = B_FALSE;
8279 * ZIL data is never be written to the L2ARC, so we don't need
8280 * special handling for its unique MAC storage.
8282 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8283 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8284 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8287 * If the data was encrypted, decrypt it now. Note that
8288 * we must check the bp here and not the hdr, since the
8289 * hdr does not have its encryption parameters updated
8290 * until arc_read_done().
8292 if (BP_IS_ENCRYPTED(bp)) {
8293 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
8295 zio_crypt_decode_params_bp(bp, salt, iv);
8296 zio_crypt_decode_mac_bp(bp, mac);
8298 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8299 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8300 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8301 hdr->b_l1hdr.b_pabd, &no_crypt);
8303 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8308 * If we actually performed decryption, replace b_pabd
8309 * with the decrypted data. Otherwise we can just throw
8310 * our decryption buffer away.
8313 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8314 arc_hdr_size(hdr), hdr);
8315 hdr->b_l1hdr.b_pabd = eabd;
8318 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8323 * If the L2ARC block was compressed, but ARC compression
8324 * is disabled we decompress the data into a new buffer and
8325 * replace the existing data.
8327 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8328 !HDR_COMPRESSION_ENABLED(hdr)) {
8329 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
8330 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8332 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8333 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8334 HDR_GET_LSIZE(hdr));
8336 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8337 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8341 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8342 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8343 arc_hdr_size(hdr), hdr);
8344 hdr->b_l1hdr.b_pabd = cabd;
8346 zio->io_size = HDR_GET_LSIZE(hdr);
8357 * A read to a cache device completed. Validate buffer contents before
8358 * handing over to the regular ARC routines.
8361 l2arc_read_done(zio_t *zio)
8364 l2arc_read_callback_t *cb = zio->io_private;
8366 kmutex_t *hash_lock;
8367 boolean_t valid_cksum;
8368 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8369 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8371 ASSERT3P(zio->io_vd, !=, NULL);
8372 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8374 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8376 ASSERT3P(cb, !=, NULL);
8377 hdr = cb->l2rcb_hdr;
8378 ASSERT3P(hdr, !=, NULL);
8380 hash_lock = HDR_LOCK(hdr);
8381 mutex_enter(hash_lock);
8382 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8385 * If the data was read into a temporary buffer,
8386 * move it and free the buffer.
8388 if (cb->l2rcb_abd != NULL) {
8389 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8390 if (zio->io_error == 0) {
8392 abd_copy(hdr->b_crypt_hdr.b_rabd,
8393 cb->l2rcb_abd, arc_hdr_size(hdr));
8395 abd_copy(hdr->b_l1hdr.b_pabd,
8396 cb->l2rcb_abd, arc_hdr_size(hdr));
8401 * The following must be done regardless of whether
8402 * there was an error:
8403 * - free the temporary buffer
8404 * - point zio to the real ARC buffer
8405 * - set zio size accordingly
8406 * These are required because zio is either re-used for
8407 * an I/O of the block in the case of the error
8408 * or the zio is passed to arc_read_done() and it
8411 abd_free(cb->l2rcb_abd);
8412 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8415 ASSERT(HDR_HAS_RABD(hdr));
8416 zio->io_abd = zio->io_orig_abd =
8417 hdr->b_crypt_hdr.b_rabd;
8419 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8420 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8424 ASSERT3P(zio->io_abd, !=, NULL);
8427 * Check this survived the L2ARC journey.
8429 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8430 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8431 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8432 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8434 valid_cksum = arc_cksum_is_equal(hdr, zio);
8437 * b_rabd will always match the data as it exists on disk if it is
8438 * being used. Therefore if we are reading into b_rabd we do not
8439 * attempt to untransform the data.
8441 if (valid_cksum && !using_rdata)
8442 tfm_error = l2arc_untransform(zio, cb);
8444 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8445 !HDR_L2_EVICTED(hdr)) {
8446 mutex_exit(hash_lock);
8447 zio->io_private = hdr;
8450 mutex_exit(hash_lock);
8452 * Buffer didn't survive caching. Increment stats and
8453 * reissue to the original storage device.
8455 if (zio->io_error != 0) {
8456 ARCSTAT_BUMP(arcstat_l2_io_error);
8458 zio->io_error = SET_ERROR(EIO);
8460 if (!valid_cksum || tfm_error != 0)
8461 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8464 * If there's no waiter, issue an async i/o to the primary
8465 * storage now. If there *is* a waiter, the caller must
8466 * issue the i/o in a context where it's OK to block.
8468 if (zio->io_waiter == NULL) {
8469 zio_t *pio = zio_unique_parent(zio);
8470 void *abd = (using_rdata) ?
8471 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8473 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8475 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
8476 abd, zio->io_size, arc_read_done,
8477 hdr, zio->io_priority, cb->l2rcb_flags,
8482 kmem_free(cb, sizeof (l2arc_read_callback_t));
8486 * This is the list priority from which the L2ARC will search for pages to
8487 * cache. This is used within loops (0..3) to cycle through lists in the
8488 * desired order. This order can have a significant effect on cache
8491 * Currently the metadata lists are hit first, MFU then MRU, followed by
8492 * the data lists. This function returns a locked list, and also returns
8495 static multilist_sublist_t *
8496 l2arc_sublist_lock(int list_num)
8498 multilist_t *ml = NULL;
8501 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8505 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8508 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8511 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8514 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8521 * Return a randomly-selected sublist. This is acceptable
8522 * because the caller feeds only a little bit of data for each
8523 * call (8MB). Subsequent calls will result in different
8524 * sublists being selected.
8526 idx = multilist_get_random_index(ml);
8527 return (multilist_sublist_lock(ml, idx));
8531 * Evict buffers from the device write hand to the distance specified in
8532 * bytes. This distance may span populated buffers, it may span nothing.
8533 * This is clearing a region on the L2ARC device ready for writing.
8534 * If the 'all' boolean is set, every buffer is evicted.
8537 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8540 arc_buf_hdr_t *hdr, *hdr_prev;
8541 kmutex_t *hash_lock;
8544 buflist = &dev->l2ad_buflist;
8546 if (!all && dev->l2ad_first) {
8548 * This is the first sweep through the device. There is
8554 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
8556 * When nearing the end of the device, evict to the end
8557 * before the device write hand jumps to the start.
8559 taddr = dev->l2ad_end;
8561 taddr = dev->l2ad_hand + distance;
8563 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8564 uint64_t, taddr, boolean_t, all);
8567 mutex_enter(&dev->l2ad_mtx);
8568 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8569 hdr_prev = list_prev(buflist, hdr);
8571 hash_lock = HDR_LOCK(hdr);
8574 * We cannot use mutex_enter or else we can deadlock
8575 * with l2arc_write_buffers (due to swapping the order
8576 * the hash lock and l2ad_mtx are taken).
8578 if (!mutex_tryenter(hash_lock)) {
8580 * Missed the hash lock. Retry.
8582 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8583 mutex_exit(&dev->l2ad_mtx);
8584 mutex_enter(hash_lock);
8585 mutex_exit(hash_lock);
8590 * A header can't be on this list if it doesn't have L2 header.
8592 ASSERT(HDR_HAS_L2HDR(hdr));
8594 /* Ensure this header has finished being written. */
8595 ASSERT(!HDR_L2_WRITING(hdr));
8596 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8598 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
8599 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8601 * We've evicted to the target address,
8602 * or the end of the device.
8604 mutex_exit(hash_lock);
8608 if (!HDR_HAS_L1HDR(hdr)) {
8609 ASSERT(!HDR_L2_READING(hdr));
8611 * This doesn't exist in the ARC. Destroy.
8612 * arc_hdr_destroy() will call list_remove()
8613 * and decrement arcstat_l2_lsize.
8615 arc_change_state(arc_anon, hdr, hash_lock);
8616 arc_hdr_destroy(hdr);
8618 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8619 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8621 * Invalidate issued or about to be issued
8622 * reads, since we may be about to write
8623 * over this location.
8625 if (HDR_L2_READING(hdr)) {
8626 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8627 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8630 arc_hdr_l2hdr_destroy(hdr);
8632 mutex_exit(hash_lock);
8634 mutex_exit(&dev->l2ad_mtx);
8638 * Handle any abd transforms that might be required for writing to the L2ARC.
8639 * If successful, this function will always return an abd with the data
8640 * transformed as it is on disk in a new abd of asize bytes.
8643 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8648 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8649 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8650 uint64_t psize = HDR_GET_PSIZE(hdr);
8651 uint64_t size = arc_hdr_size(hdr);
8652 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8653 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8654 dsl_crypto_key_t *dck = NULL;
8655 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8656 boolean_t no_crypt = B_FALSE;
8658 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8659 !HDR_COMPRESSION_ENABLED(hdr)) ||
8660 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8661 ASSERT3U(psize, <=, asize);
8664 * If this data simply needs its own buffer, we simply allocate it
8665 * and copy the data. This may be done to elimiate a depedency on a
8666 * shared buffer or to reallocate the buffer to match asize.
8668 if (HDR_HAS_RABD(hdr) && asize != psize) {
8669 ASSERT3U(asize, >=, psize);
8670 to_write = abd_alloc_for_io(asize, ismd);
8671 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8673 abd_zero_off(to_write, psize, asize - psize);
8677 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8678 !HDR_ENCRYPTED(hdr)) {
8679 ASSERT3U(size, ==, psize);
8680 to_write = abd_alloc_for_io(asize, ismd);
8681 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8683 abd_zero_off(to_write, size, asize - size);
8687 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8688 cabd = abd_alloc_for_io(asize, ismd);
8689 tmp = abd_borrow_buf(cabd, asize);
8691 psize = zio_compress_data(compress, to_write, tmp, size);
8692 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8694 bzero((char *)tmp + psize, asize - psize);
8695 psize = HDR_GET_PSIZE(hdr);
8696 abd_return_buf_copy(cabd, tmp, asize);
8700 if (HDR_ENCRYPTED(hdr)) {
8701 eabd = abd_alloc_for_io(asize, ismd);
8704 * If the dataset was disowned before the buffer
8705 * made it to this point, the key to re-encrypt
8706 * it won't be available. In this case we simply
8707 * won't write the buffer to the L2ARC.
8709 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8714 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8715 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8716 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8722 abd_copy(eabd, to_write, psize);
8725 abd_zero_off(eabd, psize, asize - psize);
8727 /* assert that the MAC we got here matches the one we saved */
8728 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8729 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8731 if (to_write == cabd)
8738 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8739 *abd_out = to_write;
8744 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8755 * Find and write ARC buffers to the L2ARC device.
8757 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8758 * for reading until they have completed writing.
8759 * The headroom_boost is an in-out parameter used to maintain headroom boost
8760 * state between calls to this function.
8762 * Returns the number of bytes actually written (which may be smaller than
8763 * the delta by which the device hand has changed due to alignment).
8766 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8768 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8769 uint64_t write_asize, write_psize, write_lsize, headroom;
8771 l2arc_write_callback_t *cb;
8773 uint64_t guid = spa_load_guid(spa);
8775 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8778 write_lsize = write_asize = write_psize = 0;
8780 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8781 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8784 * Copy buffers for L2ARC writing.
8786 for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8787 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8788 uint64_t passed_sz = 0;
8790 VERIFY3P(mls, !=, NULL);
8793 * L2ARC fast warmup.
8795 * Until the ARC is warm and starts to evict, read from the
8796 * head of the ARC lists rather than the tail.
8798 if (arc_warm == B_FALSE)
8799 hdr = multilist_sublist_head(mls);
8801 hdr = multilist_sublist_tail(mls);
8803 headroom = target_sz * l2arc_headroom;
8804 if (zfs_compressed_arc_enabled)
8805 headroom = (headroom * l2arc_headroom_boost) / 100;
8807 for (; hdr; hdr = hdr_prev) {
8808 kmutex_t *hash_lock;
8809 abd_t *to_write = NULL;
8811 if (arc_warm == B_FALSE)
8812 hdr_prev = multilist_sublist_next(mls, hdr);
8814 hdr_prev = multilist_sublist_prev(mls, hdr);
8816 hash_lock = HDR_LOCK(hdr);
8817 if (!mutex_tryenter(hash_lock)) {
8819 * Skip this buffer rather than waiting.
8824 passed_sz += HDR_GET_LSIZE(hdr);
8825 if (passed_sz > headroom) {
8829 mutex_exit(hash_lock);
8833 if (!l2arc_write_eligible(guid, hdr)) {
8834 mutex_exit(hash_lock);
8839 * We rely on the L1 portion of the header below, so
8840 * it's invalid for this header to have been evicted out
8841 * of the ghost cache, prior to being written out. The
8842 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8844 ASSERT(HDR_HAS_L1HDR(hdr));
8846 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8847 ASSERT3U(arc_hdr_size(hdr), >, 0);
8848 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8850 uint64_t psize = HDR_GET_PSIZE(hdr);
8851 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8854 if ((write_asize + asize) > target_sz) {
8856 mutex_exit(hash_lock);
8861 * We rely on the L1 portion of the header below, so
8862 * it's invalid for this header to have been evicted out
8863 * of the ghost cache, prior to being written out. The
8864 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8866 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8867 ASSERT(HDR_HAS_L1HDR(hdr));
8869 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8870 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8872 ASSERT3U(arc_hdr_size(hdr), >, 0);
8875 * If this header has b_rabd, we can use this since it
8876 * must always match the data exactly as it exists on
8877 * disk. Otherwise, the L2ARC can normally use the
8878 * hdr's data, but if we're sharing data between the
8879 * hdr and one of its bufs, L2ARC needs its own copy of
8880 * the data so that the ZIO below can't race with the
8881 * buf consumer. To ensure that this copy will be
8882 * available for the lifetime of the ZIO and be cleaned
8883 * up afterwards, we add it to the l2arc_free_on_write
8884 * queue. If we need to apply any transforms to the
8885 * data (compression, encryption) we will also need the
8888 if (HDR_HAS_RABD(hdr) && psize == asize) {
8889 to_write = hdr->b_crypt_hdr.b_rabd;
8890 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8891 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8892 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8894 to_write = hdr->b_l1hdr.b_pabd;
8897 arc_buf_contents_t type = arc_buf_type(hdr);
8899 ret = l2arc_apply_transforms(spa, hdr, asize,
8902 arc_hdr_clear_flags(hdr,
8903 ARC_FLAG_L2_WRITING);
8904 mutex_exit(hash_lock);
8908 l2arc_free_abd_on_write(to_write, asize, type);
8913 * Insert a dummy header on the buflist so
8914 * l2arc_write_done() can find where the
8915 * write buffers begin without searching.
8917 mutex_enter(&dev->l2ad_mtx);
8918 list_insert_head(&dev->l2ad_buflist, head);
8919 mutex_exit(&dev->l2ad_mtx);
8922 sizeof (l2arc_write_callback_t), KM_SLEEP);
8923 cb->l2wcb_dev = dev;
8924 cb->l2wcb_head = head;
8925 pio = zio_root(spa, l2arc_write_done, cb,
8929 hdr->b_l2hdr.b_dev = dev;
8930 hdr->b_l2hdr.b_hits = 0;
8932 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8933 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
8935 mutex_enter(&dev->l2ad_mtx);
8936 list_insert_head(&dev->l2ad_buflist, hdr);
8937 mutex_exit(&dev->l2ad_mtx);
8939 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
8940 arc_hdr_size(hdr), hdr);
8942 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8943 hdr->b_l2hdr.b_daddr, asize, to_write,
8944 ZIO_CHECKSUM_OFF, NULL, hdr,
8945 ZIO_PRIORITY_ASYNC_WRITE,
8946 ZIO_FLAG_CANFAIL, B_FALSE);
8948 write_lsize += HDR_GET_LSIZE(hdr);
8949 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8952 write_psize += psize;
8953 write_asize += asize;
8954 dev->l2ad_hand += asize;
8956 mutex_exit(hash_lock);
8958 (void) zio_nowait(wzio);
8961 multilist_sublist_unlock(mls);
8967 /* No buffers selected for writing? */
8969 ASSERT0(write_lsize);
8970 ASSERT(!HDR_HAS_L1HDR(head));
8971 kmem_cache_free(hdr_l2only_cache, head);
8975 ASSERT3U(write_asize, <=, target_sz);
8976 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8977 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8978 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8979 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8980 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8983 * Bump device hand to the device start if it is approaching the end.
8984 * l2arc_evict() will already have evicted ahead for this case.
8986 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8987 dev->l2ad_hand = dev->l2ad_start;
8988 dev->l2ad_first = B_FALSE;
8991 dev->l2ad_writing = B_TRUE;
8992 (void) zio_wait(pio);
8993 dev->l2ad_writing = B_FALSE;
8995 return (write_asize);
8999 * This thread feeds the L2ARC at regular intervals. This is the beating
9000 * heart of the L2ARC.
9004 l2arc_feed_thread(void *unused)
9009 uint64_t size, wrote;
9010 clock_t begin, next = ddi_get_lbolt();
9011 fstrans_cookie_t cookie;
9013 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9015 mutex_enter(&l2arc_feed_thr_lock);
9017 cookie = spl_fstrans_mark();
9018 while (l2arc_thread_exit == 0) {
9019 CALLB_CPR_SAFE_BEGIN(&cpr);
9020 (void) cv_timedwait_sig(&l2arc_feed_thr_cv,
9021 &l2arc_feed_thr_lock, next);
9022 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9023 next = ddi_get_lbolt() + hz;
9026 * Quick check for L2ARC devices.
9028 mutex_enter(&l2arc_dev_mtx);
9029 if (l2arc_ndev == 0) {
9030 mutex_exit(&l2arc_dev_mtx);
9033 mutex_exit(&l2arc_dev_mtx);
9034 begin = ddi_get_lbolt();
9037 * This selects the next l2arc device to write to, and in
9038 * doing so the next spa to feed from: dev->l2ad_spa. This
9039 * will return NULL if there are now no l2arc devices or if
9040 * they are all faulted.
9042 * If a device is returned, its spa's config lock is also
9043 * held to prevent device removal. l2arc_dev_get_next()
9044 * will grab and release l2arc_dev_mtx.
9046 if ((dev = l2arc_dev_get_next()) == NULL)
9049 spa = dev->l2ad_spa;
9050 ASSERT3P(spa, !=, NULL);
9053 * If the pool is read-only then force the feed thread to
9054 * sleep a little longer.
9056 if (!spa_writeable(spa)) {
9057 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9058 spa_config_exit(spa, SCL_L2ARC, dev);
9063 * Avoid contributing to memory pressure.
9065 if (arc_reclaim_needed()) {
9066 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9067 spa_config_exit(spa, SCL_L2ARC, dev);
9071 ARCSTAT_BUMP(arcstat_l2_feeds);
9073 size = l2arc_write_size();
9076 * Evict L2ARC buffers that will be overwritten.
9078 l2arc_evict(dev, size, B_FALSE);
9081 * Write ARC buffers.
9083 wrote = l2arc_write_buffers(spa, dev, size);
9086 * Calculate interval between writes.
9088 next = l2arc_write_interval(begin, size, wrote);
9089 spa_config_exit(spa, SCL_L2ARC, dev);
9091 spl_fstrans_unmark(cookie);
9093 l2arc_thread_exit = 0;
9094 cv_broadcast(&l2arc_feed_thr_cv);
9095 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9100 l2arc_vdev_present(vdev_t *vd)
9104 mutex_enter(&l2arc_dev_mtx);
9105 for (dev = list_head(l2arc_dev_list); dev != NULL;
9106 dev = list_next(l2arc_dev_list, dev)) {
9107 if (dev->l2ad_vdev == vd)
9110 mutex_exit(&l2arc_dev_mtx);
9112 return (dev != NULL);
9116 * Add a vdev for use by the L2ARC. By this point the spa has already
9117 * validated the vdev and opened it.
9120 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9122 l2arc_dev_t *adddev;
9124 ASSERT(!l2arc_vdev_present(vd));
9127 * Create a new l2arc device entry.
9129 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9130 adddev->l2ad_spa = spa;
9131 adddev->l2ad_vdev = vd;
9132 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
9133 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9134 adddev->l2ad_hand = adddev->l2ad_start;
9135 adddev->l2ad_first = B_TRUE;
9136 adddev->l2ad_writing = B_FALSE;
9137 list_link_init(&adddev->l2ad_node);
9139 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9141 * This is a list of all ARC buffers that are still valid on the
9144 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9145 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9147 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9148 zfs_refcount_create(&adddev->l2ad_alloc);
9151 * Add device to global list
9153 mutex_enter(&l2arc_dev_mtx);
9154 list_insert_head(l2arc_dev_list, adddev);
9155 atomic_inc_64(&l2arc_ndev);
9156 mutex_exit(&l2arc_dev_mtx);
9160 * Remove a vdev from the L2ARC.
9163 l2arc_remove_vdev(vdev_t *vd)
9165 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
9168 * Find the device by vdev
9170 mutex_enter(&l2arc_dev_mtx);
9171 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
9172 nextdev = list_next(l2arc_dev_list, dev);
9173 if (vd == dev->l2ad_vdev) {
9178 ASSERT3P(remdev, !=, NULL);
9181 * Remove device from global list
9183 list_remove(l2arc_dev_list, remdev);
9184 l2arc_dev_last = NULL; /* may have been invalidated */
9185 atomic_dec_64(&l2arc_ndev);
9186 mutex_exit(&l2arc_dev_mtx);
9189 * Clear all buflists and ARC references. L2ARC device flush.
9191 l2arc_evict(remdev, 0, B_TRUE);
9192 list_destroy(&remdev->l2ad_buflist);
9193 mutex_destroy(&remdev->l2ad_mtx);
9194 zfs_refcount_destroy(&remdev->l2ad_alloc);
9195 kmem_free(remdev, sizeof (l2arc_dev_t));
9201 l2arc_thread_exit = 0;
9203 l2arc_writes_sent = 0;
9204 l2arc_writes_done = 0;
9206 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9207 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9208 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9209 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9211 l2arc_dev_list = &L2ARC_dev_list;
9212 l2arc_free_on_write = &L2ARC_free_on_write;
9213 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9214 offsetof(l2arc_dev_t, l2ad_node));
9215 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9216 offsetof(l2arc_data_free_t, l2df_list_node));
9223 * This is called from dmu_fini(), which is called from spa_fini();
9224 * Because of this, we can assume that all l2arc devices have
9225 * already been removed when the pools themselves were removed.
9228 l2arc_do_free_on_write();
9230 mutex_destroy(&l2arc_feed_thr_lock);
9231 cv_destroy(&l2arc_feed_thr_cv);
9232 mutex_destroy(&l2arc_dev_mtx);
9233 mutex_destroy(&l2arc_free_on_write_mtx);
9235 list_destroy(l2arc_dev_list);
9236 list_destroy(l2arc_free_on_write);
9242 if (!(spa_mode_global & FWRITE))
9245 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9246 TS_RUN, defclsyspri);
9252 if (!(spa_mode_global & FWRITE))
9255 mutex_enter(&l2arc_feed_thr_lock);
9256 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9257 l2arc_thread_exit = 1;
9258 while (l2arc_thread_exit != 0)
9259 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9260 mutex_exit(&l2arc_feed_thr_lock);
9263 #if defined(_KERNEL)
9264 EXPORT_SYMBOL(arc_buf_size);
9265 EXPORT_SYMBOL(arc_write);
9266 EXPORT_SYMBOL(arc_read);
9267 EXPORT_SYMBOL(arc_buf_info);
9268 EXPORT_SYMBOL(arc_getbuf_func);
9269 EXPORT_SYMBOL(arc_add_prune_callback);
9270 EXPORT_SYMBOL(arc_remove_prune_callback);
9273 module_param(zfs_arc_min, ulong, 0644);
9274 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
9276 module_param(zfs_arc_max, ulong, 0644);
9277 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
9279 module_param(zfs_arc_meta_limit, ulong, 0644);
9280 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
9282 module_param(zfs_arc_meta_limit_percent, ulong, 0644);
9283 MODULE_PARM_DESC(zfs_arc_meta_limit_percent,
9284 "Percent of arc size for arc meta limit");
9286 module_param(zfs_arc_meta_min, ulong, 0644);
9287 MODULE_PARM_DESC(zfs_arc_meta_min, "Min arc metadata");
9289 module_param(zfs_arc_meta_prune, int, 0644);
9290 MODULE_PARM_DESC(zfs_arc_meta_prune, "Meta objects to scan for prune");
9292 module_param(zfs_arc_meta_adjust_restarts, int, 0644);
9293 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts,
9294 "Limit number of restarts in arc_adjust_meta");
9296 module_param(zfs_arc_meta_strategy, int, 0644);
9297 MODULE_PARM_DESC(zfs_arc_meta_strategy, "Meta reclaim strategy");
9299 module_param(zfs_arc_grow_retry, int, 0644);
9300 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
9302 module_param(zfs_arc_p_dampener_disable, int, 0644);
9303 MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener");
9305 module_param(zfs_arc_shrink_shift, int, 0644);
9306 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
9308 module_param(zfs_arc_pc_percent, uint, 0644);
9309 MODULE_PARM_DESC(zfs_arc_pc_percent,
9310 "Percent of pagecache to reclaim arc to");
9312 module_param(zfs_arc_p_min_shift, int, 0644);
9313 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
9315 module_param(zfs_arc_average_blocksize, int, 0444);
9316 MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size");
9318 module_param(zfs_compressed_arc_enabled, int, 0644);
9319 MODULE_PARM_DESC(zfs_compressed_arc_enabled, "Disable compressed arc buffers");
9321 module_param(zfs_arc_min_prefetch_ms, int, 0644);
9322 MODULE_PARM_DESC(zfs_arc_min_prefetch_ms, "Min life of prefetch block in ms");
9324 module_param(zfs_arc_min_prescient_prefetch_ms, int, 0644);
9325 MODULE_PARM_DESC(zfs_arc_min_prescient_prefetch_ms,
9326 "Min life of prescient prefetched block in ms");
9328 module_param(l2arc_write_max, ulong, 0644);
9329 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
9331 module_param(l2arc_write_boost, ulong, 0644);
9332 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
9334 module_param(l2arc_headroom, ulong, 0644);
9335 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
9337 module_param(l2arc_headroom_boost, ulong, 0644);
9338 MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
9340 module_param(l2arc_feed_secs, ulong, 0644);
9341 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
9343 module_param(l2arc_feed_min_ms, ulong, 0644);
9344 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
9346 module_param(l2arc_noprefetch, int, 0644);
9347 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
9349 module_param(l2arc_feed_again, int, 0644);
9350 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
9352 module_param(l2arc_norw, int, 0644);
9353 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");
9355 module_param(zfs_arc_lotsfree_percent, int, 0644);
9356 MODULE_PARM_DESC(zfs_arc_lotsfree_percent,
9357 "System free memory I/O throttle in bytes");
9359 module_param(zfs_arc_sys_free, ulong, 0644);
9360 MODULE_PARM_DESC(zfs_arc_sys_free, "System free memory target size in bytes");
9362 module_param(zfs_arc_dnode_limit, ulong, 0644);
9363 MODULE_PARM_DESC(zfs_arc_dnode_limit, "Minimum bytes of dnodes in arc");
9365 module_param(zfs_arc_dnode_limit_percent, ulong, 0644);
9366 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent,
9367 "Percent of ARC meta buffers for dnodes");
9369 module_param(zfs_arc_dnode_reduce_percent, ulong, 0644);
9370 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent,
9371 "Percentage of excess dnodes to try to unpin");